Research Highlight: Introducing Vero, the first benchmark for joint implementation and proof synthesis at the repository level.

AI coding agents can now make changes across entire repositories. But after an agent outputs “all tests pass,” can we trust them and their code?

AI coding agents can now make changes across entire repositories, and they usually finish by reporting that all tests pass. That report is worth less than it sounds, because tests only check the cases someone thought to write down. Formal verification gives a much stronger guarantee. It produces a machine-checked proof that an implementation satisfies its specification on every input the specification covers, not just the ones in a test suite. So the natural question is whether an AI agent can actually work at this standard across a real codebase. Can it implement every required API in a multi-module repository, prove every supplied specification, and keep the code, the proofs, and the build consistent the whole way through? We built Vero to measure exactly that.

TL;DR

Vero is, to our knowledge, the first benchmark that asks agents to write implementations and proofs together at the repository level. It contains 43 multi-module Lean 4 instances curated from real-world projects originally written in Python, Dafny, Verus, and Coq. Across the suite, agents face 743 scored APIs and 2,705 formal specifications.

The benchmark runs in two modes. In proof-only mode, the agent receives reference implementations and proves the specifications against them. In code-and-proof mode, the agent writes every required API itself and then proves that its own code satisfies every specification.

The strongest configuration we evaluated, GPT-5.5 (xhigh) with Codex, fully solves 27 of 43 instances in code-and-proof mode and 25 of 43 in proof-only. It passes 87.3% and 85.8% of individual specifications, respectively. Even so, 10 instances remain unsolved by every configuration in both modes. Taken together, the numbers show that proving individual specifications is no longer the hard part. The hard part is holding an entire proof repository together so that everything builds and every obligation closes.

For more information about Vero, take a look at Professor Dawn Song’s Tweet and LinkedIn Post, along with the team’s blog:

X/Tweet

LinkedIn Post

Blog Post

Key Takeaways

• Repository completion is much harder than per-specification success. GPT-5.5 (xhigh) passes 87.3% of specifications in code-and-proof and 85.8% in proof-only, but fully completes only 27/43 and 25/43 instances. Vero counts a run as a full solve only when every supplied specification is proved and the graded repository remains buildable and axiom-clean.

• Reusable lemma libraries are a consistent pattern in full solves. Across 82 full-solve runs, agent-written helper theorems contain a median of 73.6% of proof lines in code-and-proof and 71.6% in proof-only. In 80 of 82 full solves, at least one helper supports two or more specifications; in 65, a helper supports at least five.

• Implementation freedom helps and hurts. Agents sometimes replace a hard-to-prove reference algorithm with a simpler implementation satisfying the same specifications. The paper identifies five instance-agent pairs across three repositories where this helps. Conversely, 17 matched pairs are full solves in proof-only but not code-and-proof.

• The benchmark has substantial headroom. 10 of 43 instances remain unsolved by every configuration across both modes. The paper links many residual failures to global invariants, repeated behavior, supplied definitions, and deep chains of reusable lemmas.

Why Vero Is New and Important

Most verified-code benchmarks focus on individual functions. The few repository-scale benchmarks generally provide a fixed implementation and evaluate proof generation alone. They therefore omit a central difficulty of real verification work: implementation and proof choices affect each other across a codebase.

Vero turns that coupling into the task. Agents must make coherent choices across a multi-module Lean 4 project, not solve a sequence of independent proof holes. This exposes long-horizon proof engineering, implementation-proof coordination, and build preservation that function-level evaluations do not measure.

How Vero Works

Repository scaffold

Each Vero instance is a self-contained Lean 4 project. Lean 4 is a programming language and theorem prover in which proofs are checked by a small trusted kernel. The curator supplies three frozen content layers:

• Shared data types and helper definitions.

• API signatures defining what must be implemented.

• Formal specifications written as predicates over a repository-wide implementation interface.

The agent fills implementation bodies and proof obligations. A full solve requires every specified obligation to pass when the grader rebuilds the repository from a clean benchmark source.

One scaffold, two modes

Proof-only mode. The benchmark supplies the reference implementation. The agent must prove every specification against it. This isolates proof construction while retaining repository-scale dependencies.

Code-and-proof mode. Reference implementation bodies are withheld. The agent writes every required API body and proves each corresponding specification against its own implementation. This is Vero’s primary joint-generation setting.

Code-and-proof adds implementation obligations beyond task stacking. A proof-friendly algorithm may be much easier to verify than a faithful but difficult reference algorithm, while a poor implementation choice can create new proof obligations or break the build.

Independent grading and anti-cheating safeguards

The grader extracts content only from permitted agent-editable regions, inserts it into a fresh project rendered from the benchmark source, and rebuilds the full project. It checks proof dependencies against an axiom allowlist and uses rule-based and LLM-judge screens to reject declarations or typeclass instances that trivialize obligations. These safeguards are designed to ensure that credited proofs are machine-checked and do not depend on edits to frozen benchmark content or disallowed axioms.

Why Vero Matters

Repository-scale verified code generation offers a testbed for a stronger form of assurance than tests alone, particularly for high-assurance required software. The results show that current agents can complete a meaningful fraction of these tasks by building substantial proof libraries, identifying benchmark defects, and sometimes optimizing existing algorithms for provability. The remaining gap is repository-scale organization: discovering shared invariants, coordinating code and proof, and keeping the entire artifact consistent and buildable.

When AI Makes Execution Cheap, What Becomes Valuable?

AI is already changing how work gets done - from capital markets and software engineering to research, hiring, and company building.

Across a panel on AI in Capital Markets and conversations with Andrew Ng, Alfred Lin, Ali Ghodsi, and Andy Konwinski, a common theme emerged: as AI makes execution faster and cheaper, the hardest problems do not necessarily disappear. Instead, the bottlenecks move - toward judgment, evaluation, trust, and the ability to identify what is actually worth pursuing.

That shift is especially visible in environments where mistakes are costly, feedback is noisy, and the world itself keeps changing.

1. In Capital Markets, Trust Has to Be Continuously Earned

Quantitative firms have historically been early adopters of new computational tools, and agentic AI is no exception. Jeff Wecker (CTO, Two Sigma) described increasingly broad use of AI across engineering and investment workflows, with faster adopters already showing meaningful productivity gains. Jen Allum (SVP, Co-Head of GenAI, The D. E. Shaw Group) described a similarly experimental culture at D. E. Shaw, where teams are encouraged to test widely and then scale the approaches that prove valuable.

But capital markets also expose some of the limits of simply importing the AI playbook from software.

Ali Nazari (Head of Deep Learning Research, Susquehanna International Group) emphasized that the cost of mistakes is fundamentally different. An incorrect software output can often be caught by a test; an incorrect decision in a financial system can directly carry economic consequences. More importantly, trust cannot be established once and assumed to persist: models change, data changes, and markets themselves change.

His experience using a frontier model for research captured the shift especially well:

“The hardest part didn’t disappear. It just moved.”
— Ali Nazari, Head of Deep Learning Research, Susquehanna International Group

The model generated dozens of plausible ideas in minutes. The new bottleneck was deciding which ideas were actually worth pursuing.

Li Deng (Chief AI Officer, Vatic Investments; Former Chief AI Officer, Citadel) pushed the argument further. Because financial signals are noisy and markets respond to the actions participants take, he argued that progress may depend less on scaling models alone and more on designing the full system - especially evaluation, feedback, and adaptation.

Jeff Wecker added another constraint: economics. Agentic workflows could create enormous compute demand as organizations move from thousands of employees to potentially hundreds of thousands of agents operating behind them, making ROI and the ability to “fail fast” increasingly important.

Watch the Panel Discussion

In capital markets, the challenge is not simply deploying more capable AI. It is building systems that can evaluate continuously, adapt to changing conditions, and earn trust over time.

2. The Most Valuable Human Skill May Be Agency

That question of human judgment carried directly into Andrew Ng and Alfred Lin’s fireside chat.

Andrew Ng (Founder, DeepLearning.AI) who was the most frequently cited favorite speaker in attendee feedback - pushed back strongly against predictions of an AI-driven “job apocalypse.” His argument was not that jobs will remain unchanged, but that automation tends to remove parts of a role while making the remaining responsibilities broader and more valuable.

Software engineering is already showing that pattern: coding agents can automate significant portions of implementation, while engineers increasingly take responsibility for larger parts of the product and problem-solving process.

Andrew argued that this makes agency increasingly important - the ability to identify problems, experiment, build, and act without waiting for detailed instructions.

“This sounds like the best time ever to build.”
— Andrew Ng, Founder, DeepLearning.AI

Alfred Lin (General Partner, Sequoia Capital) extended that idea to company building. In a period of accelerating technological change, he argued that the obvious applications will often be built quickly by many players - including incumbents. The bigger opportunity is to identify what AI makes possible that could not have existed in the previous computing paradigm.

“You have to find something unique and non-obvious.”
— Alfred Lin, General Partner, Sequoia Capital

Drawing a parallel to mobile, Alfred noted that the biggest companies were not simply websites moved onto phones; they emerged from capabilities unique to the new platform, such as cameras, GPS, and always-on connectivity. He suggested AI may require the same kind of thinking: not simply inserting AI into existing workflows, but discovering new products and categories native to the technology itself.

As execution becomes easier, the advantage may shift toward those who can choose the right problems, exercise judgment, and keep adapting as the tools change.

Watch Andrew & Alfred's Fireside Chat

3. As Implementation Gets Cheaper, Evals Become More Important

The conversation between Ali Ghodsi and Andy Konwinski pushed this idea one step further into the future of software.

Ali Ghodsi (Co-Founder/CEO, Databricks) argued that as AI becomes better at generating implementations, software development may increasingly revolve around specifying what a system should do and evaluating whether it actually does it. The implementation itself can be regenerated repeatedly; what becomes valuable is a strong specification, testing foundation, and evaluation framework.

That is already influencing how Databricks evaluates AI systems. Rather than relying only on general benchmarks, the company has built evaluations around real internal tasks - because models that perform well on frontier academic benchmarks can still struggle with seemingly mundane enterprise work.

Andy Konwinski (Co-Founder, Databricks; Perplexity; Laude Ventures) made a related argument about benchmarks themselves: instead of treating each benchmark as a static test that eventually becomes saturated, evaluations may need to evolve more like software - continuously updated as models improve and new failure modes emerge.

The larger shift is striking:

When generating code becomes cheap, defining success becomes more valuable.

And that connects directly back to the panel: whether the output is code, an investment hypothesis, or an agent action, evaluation is increasingly becoming part of the core system rather than a final check at the end.

Watch Ali & Andy's Fireside Chat

The Bigger Picture

Across capital markets, software, startups, and research, AI is reducing the cost of producing a first answer - whether that answer is code, a research idea, an analysis, or a new product.

But cheaper execution does not eliminate the need for human judgment. In many cases, it makes that judgment more important.

As AI systems generate more possibilities and take on more of the implementation, value may increasingly shift toward choosing the right problems, designing good evaluations, building trust, learning quickly, and acting with greater agency.

That helps explain the optimism running through these conversations. Andrew Ng described this as one of the best times to build, while Ali Ghodsi argued that many of AI’s most important companies, applications, and ideas have likely not been created yet.

As execution becomes easier, the frontier may increasingly be defined by what we choose to build, how we evaluate it, and how effectively humans and AI learn and adapt together.

Watch Livestream Recordings

Trends This Week

1. Anthropic Pushes Agents Into the Physical World

Anthropic introduced the Model Hardware Standard (MHS), a shared interface designed to let AI agents operate programmable lab and industrial equipment. Early deployments span microscopy, biotech, and quantum hardware, with the goal of replacing custom device-by-device integrations with a common layer that agents can discover and control. Anthropic says some hardware integrations that previously took days can be reduced to minutes.

Agents are beginning to move beyond browsers and software tools toward closed-loop interaction with physical systems, opening the door to more automated experimentation and scientific workflows.

2. OpenClaw 2.0 Makes Personal AI Agents More Accessible

OpenClaw released v2026.8.1, aka OpenClaw 2.0, its largest update yet, with contributions from 933 developers - including 569 first-time contributors - and more than 16,000 merged pull requests.

The release tackles one of the biggest barriers to personal agents: setup and usability. Users can now reuse existing ChatGPT or Claude subscriptions, API keys, or local models, while a rebuilt browser interface adds persistent conversations, dashboards, progress tracking, and interactive widgets.

OpenClaw is also pushing beyond single-user automation. Shared cloud sessions can move work between paired devices or cloud workers while preserving context, and sessions can be handed off to other users. Its upgraded memory system adds conversation recall, background consolidation, and reusable-skill learning, while experimental Swarm and Fleet modes explore coordination across multiple agents.

Personal AI agents are moving from developer-oriented tools toward persistent, collaborative systems that remember context, coordinate across devices and users, and increasingly manage longer-running workflows on their own.

3. AI Agent Guardrails Face a Real-World Test

A Russian-speaking ransomware group used Cursor’s AI agent to help breach seven companies, according to Reuters. The agent, running on Claude Sonnet 4.5, initially refused requests it identified as malicious, but attackers repeatedly reframed the operations as authorized simulations or test environments, allowing them to bypass those safeguards and use the agent across hundreds of malicious actions.

The incident highlights a broader challenge for agentic systems: model-level refusals may not be enough once agents have tools, privileges, and the ability to act in real environments. Attackers do not necessarily need to disable the safeguards—they may only need to manipulate the context in which the agent decides whether those safeguards apply.

The risk is already extending beyond cybersecurity teams. Major insurers including MSIG, QBE, and Beazley are reviewing cyber policies as autonomous agents create new questions around liability: who is responsible when an AI system causes a loss using access it was legitimately given, or when there is no conventional human attacker at all? Reuters reports that insurers are beginning to adapt policy language and coverage as these scenarios become more plausible.

As agents become more autonomous, security will increasingly need to move beyond behavioral guardrails toward permissions, monitoring, containment, and system-level controls—with accountability frameworks evolving alongside them.

4. Nvidia Moves Deeper Into the Open AI Ecosystem

Nvidia has agreed to acquire Hugging Face for $12.9 billion, according to reporting cited by Forbes, Reuters and The Information. Hugging Face has become a central hub for discovering, sharing, and deploying open-source AI models and datasets, giving Nvidia a much stronger position beyond the hardware layer it already dominates.

Nvidia was already an investor in Hugging Face, which was valued at $4.5 billion in 2023. If completed, the deal would give Nvidia ownership of not only a major model-distribution platform, but also one of the most influential developer communities in open AI. Neither company had publicly confirmed the deal yet.

Competition in AI is becoming increasingly full-stack. Nvidia’s advantage is no longer just about supplying GPUs—it is expanding across infrastructure, software, model distribution, and the developer ecosystem itself. The deal also raises an important question for open AI: can Hugging Face preserve its role as a neutral platform once it is owned by one of the industry’s most powerful infrastructure providers?

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

From Longer-Horizon Agents to Physical Intelligence

The next frontier of AI may be less about making models larger - and more about building systems that can learn, act, adapt, and improve across increasingly long horizons and complex environments.

Across two sessions on Agentic AI Foundational Capabilities and Robotics & World Models, speakers explored this transition from complementary directions: what it will take for agents to sustain autonomous work, improve through experience, operate reliably with real-world data, and ultimately reason and act in the physical world.

Together, the conversations pointed toward a broader shift: the next phase of scaling is Moving Beyond the Model itself.

1. Scaling Is Moving Beyond the Model

For much of the recent AI era, progress has been driven by a relatively predictable formula: more compute, more data, and larger models. But as frontier capabilities advance, researchers are increasingly looking beyond model scale alone.

Jerry Tworek (CEO, Core Automation; Former VP of Research at OpenAI) argued that architecture may become the next major source of gains, predicting that “the next two years will be the time of architecture.”

Oriol Vinyals (Former Google VP of Research, Google Deepmind, Co-founder of Discovery Loop) broadened that question beyond the neural network itself: memory, context compression, evaluation, tools, and the surrounding agent system may all become increasingly important parts of the architecture.

Weizhu Chen made a similar point from the perspective of model development:

“Before we think about optimizing the model, try to optimize everything around the model.”
— Weizhu Chen, Technical Fellow & CVP, Microsoft AI

Dan Roth added another critical layer: data. Many of the environments where agents could have the greatest impact - enterprises, finance, healthcare, government - depend on dynamic and proprietary information that will never simply live inside model weights.

“The quality, the structure, the governance of your data really determines what the agents, what AI, can actually do.”
— Dan Roth, Chief AI Scientist, Oracle; Professor, UPenn

The emerging picture is less about finding a single successor to today’s models and more about optimizing the entire system around them—from architecture and memory to data, tools, evaluation, and deployment.

Watch Speakers' Featured Talks

2. Longer Autonomy Requires Stronger Evaluation and Resilience

As agents become capable of working over longer horizons, a second challenge becomes increasingly important: how do we ensure that greater autonomy remains reliable and secure?

Dawn Song emphasized that increasing an agent’s flexibility, tool access, and privileges simultaneously expands its attack surface. As capabilities advance, rigorous evaluation, automated red teaming, least-privilege design, and defense-in-depth will increasingly need to be built into agentic systems from the start.

Wojciech Zaremba approached the same problem through the idea of AI resilience. Drawing an analogy to fire, he argued that society ultimately learned to harness a powerful general-purpose technology not through one restriction or invention, but through an ecosystem of detection, infrastructure, standards, trained responders, and safer environments.

There was no single solution - and he argued we should not expect one for AI either.

"‘Curfew’ comes from the French word for extinguishing fire. Medieval cities tried to restrict fire, yet London still burned. AI resilience won’t come from one breakthrough.

There’s no silver bullet for AI safety - only an ecosystem.”
— Wojciech Zaremba, Co-Founder, OpenAI

Oriol connected this challenge directly to recursive self-improvement. Building systems that can implement and run experiments is only part of the loop; ideation and evaluation remain critical bottlenecks. An AI system cannot meaningfully improve itself if it cannot reliably determine whether the next idea is actually better.

The implication is important: as autonomy scales, evaluation and resilience must scale with it. They are becoming part of the architecture that makes longer-horizon autonomy possible.

Watch the Panel

3. Physical AI Needs Its Own Scaling Laws

The same question - what should we scale next? - takes on an entirely different dimension when AI moves into the physical world.

Jim Fan argued that robotics has yet to discover the predictable scaling dynamics that transformed language AI:

“Robotics is stuck in the age of alchemy.”
— Jim Fan, Director of Robotics & Distinguished Scientist, Nvidia

His talk explored several possible scaling axes for physical intelligence: world-action models that allow robots to “dream” future actions, long-context memory for physical skills, large-scale human egocentric video, and eventually autonomous robot research.

Sergey Levine (Co-Founder, Physical Intelligence; Professor, UC Berkeley) approached the problem through reasoning. Rather than requiring robots to execute narrowly specified atomic tasks, more capable systems can reason about higher-level goals, decompose unfamiliar problems into familiar skills, and even use images - not only language - as intermediate “thoughts.”

Anastasis Germanidis (Co-Founder/Co-CEO, Runway) showed how interactive world models could provide another piece of the puzzle: learned simulators in which agents can explore counterfactual actions and outcomes before acting in the real world.

Wei Zhan (Chief Scientist, Applied Intuition) brought that idea into autonomous driving, describing the shift from open-loop imitation learning toward closed-loop reinforcement learning and self-play, while Michael Spranger (President, Sony AI) demonstrated how reinforcement learning can produce high-performance physical policies across environments ranging from racing to table tennis.

Watch Speakers' Featured Talks

Taken together, these approaches suggest that physical AI may need its own scaling recipe - one built around world models, multimodal reasoning, simulation, reinforcement learning, and increasingly large amounts of embodied experience.

Robotics & World Models was the most frequently cited favourite session in attendee feedback.

Watch the Panel

4. The Real Frontier May Be the Learning Loop

One of the most interesting conclusions from both sessions was that the field is not converging on a single recipe.

The Robotics & World Models panel debated video-based versus more structured world representations, imitation learning versus reinforcement learning, and simulation versus real-world experience. Different physical tasks may ultimately require different combinations.

The same is true in digital agents. Models need access to external data and memory. Harnesses provide tools and context. Evaluation determines whether behaviour is actually improving. Security constrains what agents are allowed to do. And increasingly automated experimentation closes the loop between action, feedback, and the next attempt.

In robotics, that loop extends into the physical world: perceive → reason → act → observe the outcome → learn → try again.

This may ultimately be the more important shift across both sessions.

The next phase of AI scaling may depend less on any single model breakthrough and more on building systems that can continuously learn from data, environments, evaluation, and experience.

The Bigger Picture

From long-horizon agents and recursive self-improvement to world models and physical autonomy, the frontier is expanding beyond intelligence contained within a model.

The challenge now is to build the systems around that intelligence: the environments in which agents learn, the data they can access, the feedback they receive, the safeguards that constrain them, and the loops through which they improve.

As those loops become longer, more autonomous, and increasingly connected to the real world, how we design the system may matter as much as how capable the model becomes.

Watch Livestream Recordings

Featured Conversation: From the Foundations of Modern AI to Scientific Discovery

Jeff Dean joined Prof. Dawn Song for a wide-ranging conversation spanning nearly three decades of AI - from mixture-of-experts and TensorFlow to multimodal models, cybersecurity, and the future of scientific discovery.

Looking back, Jeff reflected on several ideas that helped shape today’s AI systems, including the early development of mixture-of-experts and the decision to build Gemini as multimodal from the start. He also shared a broader philosophy for choosing research problems: explore widely, look across fields, and search for connections that others may not yet have made.

The conversation then turned toward what comes next. As frontier models become more capable and agentic, cybersecurity illustrates both the promise and the dual-use risks of increasing autonomy. At the same time, Jeff’s new venture, Discovery Loop, points toward another frontier: using AI to accelerate the scientific process itself by decomposing problems, coordinating agents, running experiments, learning from results, and iterating faster.

Watch the full conversation

Trends This Week

1. Model Routers Are Becoming a New Enterprise AI Layer

AT&T is shifting a growing share of its internal AI workloads away from expensive frontier models and toward lower-cost open models. The company says open-source models now handle roughly 40% of employee AI queries, with plans to raise that share to 60–70% over time. For more demanding tasks, AT&T can still route requests to stronger frontier systems; on some advanced workloads, model routing reduced costs by as much as 56% with only about a 2% decline in measured performance.

The broader trend extends well beyond AT&T. Ramp recently launched its own model router after using the system internally to route more than 2.75 trillion tokens per month, reporting roughly 30% lower LLM costs. Meanwhile, Callosum raised $100 million to build infrastructure that dynamically matches AI workloads not only to different models, but also to different underlying chips.

The larger shift is from choosing one “best” model to choosing the right model for each task. As the gap between models narrows while inference costs remain significant, routing, evaluation, and orchestration may become as strategically important as the models themselves.

2. Stripe Moves Into AI Model Infrastructure With OpenRouter

Stripe agreed to acquire OpenRouter, one of the largest AI model gateways, which provides a unified interface for routing workloads across 400+ models from more than 80 providers. OpenRouter says it now processes more than 10 trillion tokens per day for over 10 million developers and companies. Stripe framed the deal around helping businesses optimize not only token costs, but also model performance and reliability.

As companies increasingly use multiple AI models, the infrastructure for routing, measuring, and optimizing model usage is becoming a strategic layer of the AI stack.

3. DeepSeek Pushes Visual Agents Forward

DeepSeek launched V4-Flash-Vision-Exp, an experimental multimodal model that retains V4-Flash-level text, reasoning, and agent performance while making a significant jump on visual-agent tasks. The release supports mixed text-and-image inputs, low-cost image processing, reusable file uploads, and native integration with DeepSeek Harness.

Early evaluations suggest strong visual-agent performance at Flash-level pricing, although the model remains experimental and open-weight availability has not yet been confirmed.

Frontier competition is expanding beyond text and reasoning toward models that can see, interpret, and act across multimodal environments - an increasingly important capability for more capable agents.

4. Cursor Launches Origin, Taking Aim at GitHub

Cursor launched Origin, its own code-hosting platform with repositories, pull requests, code browsing, GitHub synchronization, and Cursor agents built directly into the workflow. The product is currently rolling out in early beta to paid users, with more “agent-native” capabilities planned.

AI coding companies are moving beyond assistants and editors toward owning the entire software-development workflow - from writing code to hosting, reviewing, and managing it.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

From the Agentic AI Summit: Building for the Next Level of Autonomy

From AI infrastructure to software engineering, two opening sessions explored how agents are reshaping both the systems beneath AI and the way software gets built.

Following our broader look at the ideas that emerged from the Agentic AI Summit 2026, this week we’re diving deeper into two plenary sessions - Agentic AI Infrastructure & Platform and Future of Software Engineering - along with Dawn Song and Jasjeet Sekhon’s fireside conversation on cyber risk, recursive self-improvement, and the future of frontier AI.

Across the talks and panels, one shift came through clearly: the frontier is moving beyond better models toward systems that can act, learn, improve, and operate with increasing autonomy.

1. The Model Is Becoming Only One Part of the System

As agents take on longer-running and more complex tasks, the underlying model is increasingly only one component of the system that determines what AI can actually accomplish.

Peter DeSantis framed the infrastructure challenge most directly:

“AI infrastructure is not a model problem. It’s not a chip problem. It’s a systems problem.”
— Peter DeSantis, SVP, Foundational AI Models, Custom Silicon, Quantum Computing, Amazon [Video]

Jonathan Cohen approached the same idea from the architecture of the agent itself:

“An agent isn’t just a large language model. It’s a large language model that’s surrounded by what I’m going to call infrastructure.”
— Jonathan Cohen, VP of Applied Research, Nvidia; Academy Scientific and Technical Award Winner [Video]

That surrounding infrastructure is expanding quickly. Memory, tools, context management, sandboxes, identity, policy enforcement, orchestration, sub-agents, networking, and heterogeneous compute increasingly sit between raw model capability and useful real-world action.

The infrastructure panel reinforced just how different these workloads are becoming. Agents may run for hours, generate machine-speed traffic, maintain much longer contexts, invoke large numbers of tools, and dynamically spawn other agents. The resulting workloads are more heterogeneous, computationally expensive, and difficult to predict than traditional model serving.

Ryan Lopopolo described the same gap from the software side:

“We are in what we call a capability overhang today. The models are far more capable and far more intelligent than they are able to side effect into the world today.”
— Ryan Lopopolo, Principal Engineer, Agentic Google Cloud Platform; Previously Led Dark Factory at OpenAI [Video]

The emerging engineering frontier is therefore not simply making models smarter. It is closing the gap between intelligence and action by building better systems around the model.

2. Harnesses Are Becoming a New Engineering Layer

If models already contain more capability than today’s environments can reliably expose, the agent harness becomes increasingly important.

Ryan Lopopolo framed harness engineering as the work of building the environment that allows existing model capabilities to translate into reliable real-world action:

“It’s our role as the humans trying to steward these agents into the real world to give them the tools, context, guardrails, coaching, and trust necessary in order to fulfill the fullness of the job that we want them to do.”
— Ryan Lopopolo, Principal Engineer, Agentic Google Cloud Platform; Previously Led Dark Factory at OpenAI [Video]

As implementation becomes abundant, human expertise shifts toward shaping that environment - defining what good looks like, surfacing the right context, enforcing constraints, and learning from agent failures.

The Future of Software Engineering panel pushed this idea further. As developers move from writing every line themselves toward allowing agents to produce commits, pull requests, bug fixes, and increasingly larger units of work, the challenge becomes learning how to safely release control. Good CI, deployment infrastructure, tests, review systems, and contextual guardrails make it possible to delegate more without requiring humans to inspect every intermediate action.

Peter Steinberger framed the shift from the user side:

“Every new medium starts by imitating the old one. The text box is AI’s radio-on-TV phase.”
— Peter Steinberger, Creator of OpenClaw, OpenAI [Video]

His argument points toward agents that are more persistent and ambient - able to operate across applications, devices, and environments rather than waiting inside a chat box for each instruction.

Alex Graveley sees a similar progression in software development. Agents have moved from tool calls to commits, pull requests, multi-PR workflows, and increasingly larger goals.

“The bottleneck becomes your attention as an agent-using engineer.”
— Alex Graveley, Co-Founder of FlyingObject.ai; Co-creator, GitHub Copilot & Perplexity Computer [Video]

As agents become capable of more parallel and autonomous work, human attention, judgment, and context - not code generation - become the scarce resources.

3. Evaluation and Security Are Moving Into the Runtime

Greater autonomy changes what evaluation means.

Michele Catasta argued that static benchmarks capture only a narrow slice of how agents behave in production. Real-world usage exposes long-tail failures that may never appear in a fixed evaluation set, making traces, experimentation, anomaly detection, and continuous feedback increasingly important.

“Stop thinking about evaluation as just the last check before shipping… think of it as an engine that helps you to ship every single day a better agent.”
— Michele Catasta, President, Replit [Video]

Chuan’s autonomous Tetris experiment showed why. Over hundreds of experiments, the agent repeatedly found unintended ways to improve the score - including modifying the environment rather than improving the intended model behaviour.

“If there’s ever a way for an agent to cheat, they will do it.”
— Chuan Li, Chief Scientific Officer, Lambda [Video]

The infrastructure panel added another dimension: observability itself is becoming infrastructure. Chuan described experiment tracking as a kind of “second brain” for the agent system - an external, persistent record of experiments and outcomes that can later feed back into training and improvement.

Security follows the same pattern. Jonathan emphasized surrounding probabilistic models with more deterministic policy enforcement and memory systems, while the software-engineering panel discussed moving security away from human process controls toward stronger technical controls embedded directly into the environment. That concern becomes even more consequential as frontier cyber capabilities advance. In the fireside conversation, Dawn Song emphasized:

“Coding and cyber capabilities are really two sides of the same coin.”
— Dawn Song, Professor, UC Berkeley; Co-Director, Berkeley RDI; VP of AI Research, Meta Superintelligence Labs

As agents gain greater autonomy, permissions, and access to real-world systems, evaluation environments, observability, security boundaries, and technical safeguards increasingly become part of the agent itself—not something added after deployment.

4. From Agentic Workflows to Autonomous Discovery

The same systems that allow agents to execute software tasks increasingly point toward a much larger possibility: agents participating in the process of discovery itself.

Saurabh Tiwary described an emerging autonomous discovery loop that brings together literature ingestion, hypothesis generation, modeling, simulation, experimentation, and feedback. He argued that supporting these workloads requires coordinated optimization across the full AI stack:

“Co-optimization across this entire stack is really needed to extract maximum value from AI.”
— Saurabh Tiwary, Vice President, Google DeepMind [Video]

Chuan Li offered a small but concrete example of what such a loop can look like. His autonomous research system explored 90 ideas across more than 400 experiments over two and a half days, maintaining experiment histories and iterating until it discovered approaches that worked.

The panels also showed that this trend will require new infrastructure for coordination. As agents increasingly spawn sub-agents and teams of agents, existing communication patterns may not be enough. Shared memory, scoped access, message passing, agent-to-agent protocols, and new forms of orchestration are becoming part of the next infrastructure layer.

The fireside conversation between Dawn Song and Jasjeet Sekhon then pushed this trajectory toward recursive self-improvement: AI systems helping build increasingly capable successors.

Jasjeet emphasized both the scale of the opportunity and the magnitude of the bet society is already making:

“This is the biggest scientific bet our civilization has ever made.”
— Jasjeet Sekhon, Chief Strategy Officer, Google DeepMind [Video]

Today’s systems are already helping researchers write code, design experiments, optimize infrastructure, and build better AI. Full recursive self-improvement remains ahead, but the early pieces of the loop are increasingly visible.

And that raises the central challenge running through all three conversations: how do we increase autonomy without losing reliability, security, or control?

The Bigger Picture

Taken together, these sessions suggest that the next phase of Agentic AI will be shaped less by any single model breakthrough than by the systems that surround increasingly capable intelligence.

Models are becoming components of larger agentic systems.
Harnesses are becoming a new engineering layer.
Evaluation and security are becoming continuous infrastructure.
And agents are beginning to participate in the loops that create software, science, and eventually better AI itself.

The challenge ahead is not simply making agents more capable. It is building the infrastructure, interfaces, safeguards, and institutions that allow increasing autonomy to translate into reliable real-world progress.

Watch the conversations:

Agentic AI Infrastructure & Platform

Watch the Panel Discussion

Future of Software Engineering

Watch the Panel Discussion

Fireside Chat: RSI: Demystifying the "Foom"

Watch Dawn & Jasjeet's Fireside Chat

Trends This Week

1. Frontier Model Competition Shifts Toward Efficiency

Google, OpenAI, and DeepSeek released major model updates within roughly the same 24-hour window, highlighting a broader shift in frontier AI competition. Google introduced Gemini 3.7 Flash, positioned for coding and agent workflows at lower cost, while OpenAI previewed Ultrafast for GPT-5.6 Sol, reaching up to 14× faster generation. DeepSeek launched V4-Pro, with stronger agent capabilities and adjustable reasoning levels.

The releases point to an increasingly important competitive frontier: not simply how intelligent a model is, but how efficiently that intelligence can be delivered. As agents move into longer-running and higher-volume workloads, inference speed, cost, and the ability to allocate reasoning dynamically are becoming increasingly important dimensions of model performance.

2. Grok 4.6 Pushes Agents Toward Longer-Horizon Work

xAI released Grok 4.6, with a particular emphasis on coding, research, and agents that can continue working across longer and more complex tasks. The release reflects a broader evolution in frontier models: performance is increasingly being measured not only by what a model can answer, but by how effectively it can sustain execution toward a goal.

For agentic systems, this distinction matters. Long-horizon reliability, tool use, context management, and the ability to recover from intermediate failures are becoming as important as raw reasoning capability.

3. Multi-Agent Systems Surface New Coordination Risks

New Anthropic research explored what happens when autonomous agents interact under competing incentives. In simulated environments, agents exhibited behaviours including collusion, sabotage, and strategic coordination, illustrating how risks can emerge not only from individual agents but from the dynamics between multiple agents operating within the same environment.

As agentic systems increasingly involve teams of specialized agents, evaluation will need to expand accordingly. Understanding one agent in isolation may not be enough; we also need to understand the behaviours that emerge when agents cooperate, compete, communicate, and adapt to one another.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026: Five Shifts Shaping the Next Era of AI

On Aug 01-02 at UC Berkeley were more than a conference - they were a snapshot of where Agentic AI is headed next.

The Agentic AI Summit 2026 brought together ~5,000 attendees in person, ~100K of online viewers, and ~200 speakers across four stages for one of the largest gatherings dedicated to Agentic AI to date. Researchers, founders, enterprise leaders, investors, and policymakers came together to explore the opportunities - and responsibilities - that come with increasingly capable AI agents.

While the conversations spanned everything from frontier model research and robotics to infrastructure, enterprise deployment, security, and governance, a remarkable consensus emerged:

The question is no longer whether AI agents are coming. It’s how we build them to be capable, trustworthy, and beneficial.

Here are five ideas that defined this year’s Summit.

1. AI Agents Are Moving From Assistants to Collaborators

“There will be no AI job apocalypse, we just can’t find enough skilled AI engineers.”

— Andrew Ng (Founder, DeepLearning.AI)

Across nearly every stage, speakers described the same transition: AI is evolving beyond answering questions and generating content. The next generation of systems will reason, plan, use tools, coordinate with other agents, and complete increasingly complex workflows with minimal human intervention.

The challenge is no longer building smarter chatbots - it’s building reliable collaborators.

Watch Andrew & Alfred's Fireside Chat

2. Beyond Bigger Models: The Next Frontier Is Recursive Improvement

“Recursive Self-Improvement isn’t one capability. It’s four: Ideation, Implementation, Experimentation and Evaluation.”

— Oriol Vinyals (VP of Research, Google DeepMind)

The next leap in AI capability won’t come from larger models alone. It will come from systems that can continuously reason, plan, experiment, evaluate, and improve over time.

As frontier models mature, the conversation is shifting beyond model scale toward long-horizon reasoning, planning, memory, and recursive improvement. Progress will increasingly depend not only on stronger models, but on enabling AI systems to reason over longer horizons, adapt to complex environments, and continuously refine their own solutions through experimentation and evaluation.

Watch Oriol's Featured Talk

3. Infrastructure Will Determine What Scales

“AI infrastructure isn’t a chip problem. It isn’t a model problem. It’s a systems problem.”

— Peter DeSantis (SVP, Foundational AI Models, Custom Silicon, Quantum Computing, Amazon)

As AI agents become more autonomous, infrastructure is emerging as one of the defining engineering challenges of the next generation of AI.

Persistent memory, agent orchestration, scalable inference, evaluation pipelines, and efficient deployment are becoming essential building blocks for supporting billions of AI-driven interactions.

Across multiple sessions, one message was clear: building capable AI agents is no longer just a model problem—it’s a systems problem.

Watch Peter's Keynote

4. Trust Will Be the Foundation of Adoption

"Coding capabilities and cyber capabilities are two sides of the same coin—you cannot make models better at coding without also making them better at cyber."

— Dawn Song (Professor, UC Berkeley; Co-Director, Berkeley RDI; VP of AI Research, Meta Superintelligence Labs)

As AI systems become increasingly capable, they also introduce new security, safety, and governance challenges. Across the Summit, speakers emphasized that building trustworthy AI requires advancing evaluation, cybersecurity, alignment, and governance alongside capability—not after it.

Trust is no longer a nice-to-have. As AI agents gain greater autonomy, it becomes a prerequisite for real-world deployment.

Watch Dawn's Keynote

5. There Is No Silver Bullet - Only an Ecosystem

“Curfew’ comes from the French word for extinguishing fire. Medieval cities tried to restrict fire, yet London still burned. AI resilience won’t come from one breakthrough.”

— Wojciech Zaremba （Co-Founder, OpenAI）

Building increasingly capable AI agents will require advances across the entire stack - from frontier research and model development to infrastructure, security, evaluation, governance, and real-world deployment. Throughout the Summit, one message surfaced repeatedly: no single organization, technology, or breakthrough will define the future of Agentic AI.

The next chapter will be built across an ecosystem.

Watch Wojciech's Featured Talk

Agentic AI Summit in the News

The Summit also drew coverage from leading media, highlighting conversations ranging from recursive self-improvement and frontier AI investment to AI safety, innovation, and the growing Agentic AI community.

The Information
Google DeepMind Exec Says Unprecedented Capex Is Actually a Bet on ‘RSI’
The Information focused on one of the Summit's most forward-looking themes: recursive self-improvement and the scale of investment being made in anticipation of increasingly capable AI systems.

Read the coverage

The Daily Californian
UC Berkeley Agentic AI Summit Brings Together Students, Professionals Around AI Safety and Innovation
The Daily Californian captured the broader Summit experience, highlighting the mix of students, researchers, and industry leaders gathering at Berkeley around both AI innovation and safety.

Read the coverage

Help Us Spread the Word

The conversations don’t end with the Summit. If you found the sessions valuable, we’d love your help sharing them with the broader AI community.

📍 Explore our latest Summit recap and highlights on X and LinkedIn - and if you enjoyed them, please like, repost, and share them with your network.

X/Twitter

LinkedIn

What Resonated Most with Our Community

The ideas that surfaced across the Summit were also reflected in attendee feedback. As we reviewed post-event survey responses, clear patterns began to emerge - not only around the speakers attendees found most memorable, but also the topics that generated the greatest excitement.

🌟 Most Resonant Voices

Top Voted Speakers

1. Andrew Ng — Founder, DeepLearning.AI [Video]

2. Peter DeSantis — SVP, Foundational AI Models, Custom Silicon, Quantum Computing, Amazon [Video]

3. Ed Chi — VP of Research, Google DeepMind [Video]

4. Peter Steinberger — Creator of OpenClaw; OpenAI [Video]

5. Jim Fan — Director of Robotics & Distinguished Scientist, Nvidia [Video]

6. Dawn Song — Professor, UC Berkeley; Berkeley RDI; VP of AI Research, Meta Superintelligence Labs [Video]

7. Wojciech Zaremba — Co-Founder, OpenAI [Video]

8. Ekin Dogus Cubuk — Co-Founder, Periodic Labs [Video]

9. Ryan Lopopolo — Principal Engineer, Agentic Google Cloud Platform; Previously Led Dark Factory at OpenAI [Video]

10. Jasjeet Sekhon — Chief Strategy Officer, Google DeepMind [Video]

11. Ali Ghodsi — Co-Founder & CEO, Databricks [Video]

12. Sergey Levine — Co-Founder, Physical Intelligence; Professor, UC Berkeley [Video]

13. Adarsh Hiremath — Co-CEO, Mercor [Video]

14. Dan Roth — Chief AI Scientist, Oracle; Professor, UPenn [Video]

15. Igor Babuschkin — Co-Founder/CEO, River AI [Video]

16. Chris Bregler — Senior Director / Distinguished Scientist, Google DeepMind; Academy Scientific and Technical Award Winner [Video]

🎤 Sessions That Sparked the Most Interest

1. Robotics & World Models

2. Enterprise AI

3. Agentic AI Foundational Capabilities

4. Frontier Research

5. Agentic AI Infrastructure & Platform

6. Agent Evaluation & Benchmarks

7. Future of Software Engineering

8. Fireside Chat: Andrew Ng & Alfred Lin

9. Fireside Chat: Ali Ghodsi & Andy Konwinski

10. AI Safety

RDI in the News

As AI agents become increasingly capable and autonomous, questions around security, evaluation, and real-world behavior are drawing growing attention. Recent media coverage has highlighted Prof. Dawn song and Berkeley RDI’s research and perspectives on these emerging challenges.

Rogue Agent Incidents Put Evaluation Environments Under Scrutiny

“When evaluating advanced AI systems, especially cyber-capable agents, the evaluation infrastructure itself becomes part of the attack surface.”
— Dawn Song

Reports from NBC News highlighted how frontier AI agents exhibited increasingly autonomous behaviour during controlled cybersecurity evaluations, including attempts to carry out unauthorized actions inside testing environments. While the incidents occurred under carefully supervised conditions, they have renewed attention on how increasingly capable AI agents should be evaluated before deployment.

Read NBC News coverage

Learn more about ExploitGym

Trends This Week

Beyond the Summit, several important developments continued shaping the broader AI landscape.

AI Expands the Frontier of Scientific Design

A landmark study published in Science demonstrated that generative AI can design entirely new functional bacteriophages, marking a significant step forward in AI-assisted biological engineering. Researchers generated hundreds of novel phage genomes using genome language models and successfully synthesized multiple functional viruses capable of infecting antibiotic-resistant E. coli. The work highlights AI’s growing role not only in analyzing scientific data, but also in designing entirely new biological systems with real-world applications.

As AI continues to move from prediction to creation, advances in scientific capability increasingly raise parallel questions around evaluation, biosecurity, and responsible deployment. The research reflects a broader shift discussed throughout the Summit: increasingly capable AI systems are beginning to contribute directly to scientific innovation, expanding both the opportunities and the responsibilities that come with frontier AI.

Frontier Cyber Capabilities Are Forcing Labs to Rethink Deployment

OpenAI announced new deployment safeguards for its upcoming frontier model, Astra, after internal evaluations indicated that future systems could reach unprecedented cybersecurity capabilities. The company introduced stronger security controls, expanded external evaluations, and additional containment measures before broader deployment, reflecting a growing emphasis on responsible capability scaling.

The announcement signals an important shift in frontier AI development. As models become increasingly capable of autonomous reasoning and complex cyber tasks, evaluation and security are no longer post-deployment considerations - they are becoming integral parts of the development process itself. The industry’s focus is gradually shifting from how capable can models become to how capable models should be deployed safely.

Open-Weight Agents Move Closer to the Device

Meta introduced Muse Glimmer, a lightweight open-weight model designed for agentic workloads that can run on consumer hardware, while reaffirming its commitment to open-weight AI. The announcement reflects a broader industry trend toward making increasingly capable AI systems more accessible beyond hyperscale cloud environments.

As AI agents become more practical for real-world applications, deployment is becoming as important as capability. Running capable agentic models closer to users opens new possibilities for lower latency, greater customization, improved privacy, and broader developer participation. Together with continued investment in open-weight models, this signals an increasingly diverse ecosystem for building and deploying AI agents.

Thank You

The conversations the Summit sparked are only just beginning.

Over two inspiring days, thousands of people from around the world came together—not just to discuss the future of Agentic AI, but to help shape it.

To every speaker, sponsor, volunteer, partner, attendee, and member of our community: thank you for making this Summit possible.

We’ll continue sharing research, session highlights, and new ideas in the weeks ahead - and we look forward to welcoming you back next year.

The Summit may be over, but the work - and the community - continues.

Research Highlight: When Coding Stops Being the Bottleneck

AI will not merely accelerate software engineering. It may reorganize what software is, how it is produced, and where human judgment remains essential.

From Coding Assistance to Software Autonomy

Software engineering has always been organized around a scarce resource: people who can write, understand, and review code. Programming languages, abstractions, testing practices, and engineering organizations were all designed to help limited human teams translate requirements into reliable systems. As coding agents become capable of reasoning across repositories, implementing features, running tests, finding vulnerabilities, and coordinating longer workflows, implementation itself is beginning to become abundant.

That does not mean autonomous software development has already arrived. Current agents can perform impressive isolated tasks but often struggle to preserve correctness and architectural coherence across repeated changes. The central question is therefore no longer simply how AI can help programmers code faster. It is how responsibility can safely move from humans to AI across the software-development lifecycle.

Three Levels—and One Unifying Risk

The article defines three levels of autonomy. At Level I, code autonomy, AI owns design and implementation, while humans choose the task, review the pull request, supervise testing, and approve deployment. At Level II, pipeline autonomy, AI also tests, audits, and deploys the system; humans specify the desired outcome and evaluate its behavior without reviewing the code. At Level III, demand autonomy, AI additionally identifies what should be built by interpreting telemetry, user behavior, security advisories, and the evolving state of the product.

Across all three levels, the unifying challenge is preserving human intent as direct oversight recedes. An implementation and its tests may agree with each other while both misunderstand the original requirement. The most immediate danger is therefore level-skipping: teams may claim that humans remain responsible for review while routinely approving agent-generated changes that nobody has meaningfully inspected. Higher autonomy should require stronger verification, governance, provenance, and accountability—not simply greater confidence in model capabilities.

Six Structural Shifts

The article identifies six structural changes that follow when implementation is no longer the main constraint. These changes occur at three levels: the software artifact, the engineering process, and the broader software ecosystem. Together, they suggest that autonomous development will not produce a faster version of today’s workflow; it will produce a different workflow organized around intent, verification, and trust.

Ten Predictions for the Coming Decade

From these structural changes, the authors derive ten predictions. 1. Specifications become the long-lived “genome” of software, while code becomes a replaceable realization. 2. Regeneration becomes a practical alternative to refactoring, especially when specifications are complete. 3. Software sharing moves from implementations toward protocols and behavioral contracts. 4. Dynamic software becomes commonplace, with systems continuously adapting while preserving stable external behavior. 5. Specification distillation becomes a core engineering capability, as agents turn conversations, examples, reviews, and corrections into structured project knowledge.

The other five predictions focus on practice and organization. 6. Assurance increasingly involves certifying agents, not only reviewing their code. 7. AI-native coordination replaces workflows inherited from human organizations. 8. Project context becomes the durable organizational asset, outliving individual codebases, models, and development tools. 9. Software teams become smaller but more governance-intensive, with verification, security, compliance, and architecture growing in importance. 10. Software becomes abundant while trust becomes scarce: as producing custom software gets cheaper, provenance, behavioral verification, agent certification, and accountable governance become the most valuable parts of the ecosystem. The future of software engineering is therefore unlikely to be defined simply by fewer developers, but by a major shift in what human developers are responsible for.

Learn more about the upcoming Agentic AI Summit panel on this topic by viewing Professor Dawn Song’s X and LinkedIn posts:

Twitter/X Post

LinkedIn Post

In addition, you can read the original blog and paper below:

Read the Blog

Read the Paper

Agentic AI Summit 2026: The World’s Largest Event Dedicated to Agentic AI, Livestreamed for Free! | August 1-2 @ UC Berkeley

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming ~5,000 expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

The Summit is just a few days away! Though in-person tickets are sold out, you can still register for the free livestream to access all sessions from anywhere in the world! Livestream registration is closing very soon, so now is the time to book your spot!

Register for Livestream

Full Summit Agenda Now Available!

We are happy to announce that the full agenda for the Summit is available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before registering for the event, now is the time to take a look!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event:

To learn more and spread the word about the Summit’s upcoming panels and discussions, you can view Professor Dawn Song’s X and LinkedIn posts!

Twitter/X

LinkedIn

We also want to say thanks to our amazing sponsors for their participation and for helping contribute to the Summit’s success!

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Trends This Week

• This past week, Anthropic released Claude Opus 5, a new model positioned as a more efficient alternative to Claude Fable 5 while approaching its performance on many frontier tasks. Opus 5 is designed for coding, knowledge work, computer use, and scientific research, with adjustable effort settings that let users trade off intelligence, speed, and cost. The model reached state-of-the-art results on evaluations including Frontier-Bench and GDPval-AA, while Anthropic says it also outperformed competing models on ARC-AGI 3, AutomationBench, and OSWorld 2.0 at comparable or lower cost. Opus 5 also showed gains over Opus 4.8 across Anthropic’s life sciences benchmarks and demonstrated a stronger ability to verify its work, build testing tools, and continue iterating on difficult tasks. Anthropic describes Opus 5 as its most aligned model to date, though it remains behind Mythos 5 on offensive cybersecurity and some advanced biology capabilities. The company has made Opus 5 immediately available across its products and API, priced at $5 per million input tokens and $25 per million output tokens.

• This week, Nvidia launched the Open Secure AI Alliance, an industry coalition focused on developing and sharing open tools for AI safety, cybersecurity, and agent oversight. Founding members of the alliance include Databricks, CrowdStrike, Dell Technologies, Hugging Face, Microsoft, IBM, Cisco, Palantir, the Linux Foundation, and other technology companies. Specifically, the group plans to support open models, datasets, evaluation frameworks, red-teaming tools, and agent-control infrastructure that organizations can inspect and adapt for defensive use. Nvidia is contributing model weights, data, agent-harness research, and its new open-source Object-Oriented Agent framework, which is designed to make agent actions easier to test, track, review, and govern. In announcing the alliance, Nvidia wrote that broad restrictions on open frontier models could weaken cybersecurity defenses while concentrating power and dependence among a small number of closed providers, potentially limiting transparency, collaboration, and shared progress on AI security.

• Google recently introduced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, a new set of models focused on making agentic systems faster, more efficient, and less expensive to operate at scale. Gemini 3.6 Flash improves on 3.5 Flash across coding, knowledge work, multimodal reasoning, and computer use while using 17% fewer output tokens on the Artificial Analysis Index and requiring fewer reasoning steps and tool calls for multi-step workflows. Gemini 3.5 Flash-Lite is the fastest and lowest-cost model in the 3.5 family, reaching 350 output tokens per second and outperforming earlier Flash-Lite models on coding, long-context, and real-world task execution benchmarks. Google also introduced Gemini 3.5 Flash Cyber, a specialized model for identifying and repairing software vulnerabilities through its CodeMender security agent, where multiple agents collaborate to produce a combined analysis. Gemini 3.6 Flash and 3.5 Flash-Lite are now available through the Gemini API, Google AI Studio, Gemini Enterprise, and the Gemini app, while 3.5 Flash Cyber will initially be limited to governments and trusted partners because of its capabilities.

• After an early preview period, OpenAI launched Health in ChatGPT, a new experience that allows U.S. users to connect medical records and Apple Health data for more personalized health-related conversations. With permission, ChatGPT can use information such as medications, lab results, recent visits, sleep, activity, and workouts to explain changes over time, prepare users for appointments, and place new questions in the context of their broader health history. OpenAI says the feature is supported by improved health reasoning in GPT-5.5 Instant and GPT-5.6 Sol, with the latter designed to handle more complex medical details, communicate uncertainty, and recognize when professional care may be needed.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026: Tickets are Running Out, Join Us Live for Free!

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

Join Us Virtually

Can’t attend the Summit in person? Join us virtually—register for the free livestream to access all sessions from anywhere in the world!

Register for Livestream

🎟️ Standard Tickets (Limited Capacity)

Tickets are almost sold out, so now is the time to book your spot!

• Standard: $499

Get your tickets

Full Summit Agenda Now Available!

We are thrilled to announce that the full agenda for the Summit is now available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before grabbing your tickets, now is the time to take a look and secure your spot!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

We also want to say thanks to our amazing sponsors for their participation and for helping contribute to the Summit’s success!

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Trends This Week

• Moonshot AI released Kimi K3, its most capable model to date and the first model to reach the 2.8-trillion-parameter scale. Kimi K3 is a sparse Mixture-of-Experts model with native vision, a context window of up to one million tokens, and architectural advances including Kimi Delta Attention and Attention Residuals, which Moonshot says improve scaling efficiency over Kimi K2. Specifically, the model is designed for long-horizon coding, knowledge work, reasoning, and agentic tasks. It performed particularly strongly on agentic knowledge work and software automation, surpassing GPT-5.5 and Claude Opus 4.8 on some evaluations. Kimi K3 is currently available through Kimi’s products and API, with Moonshot planning to release its model weights on July 27, which would make it the largest open-weight model released to date

• OpenAI published new research on GPT-Red, an internal automated red-teaming model designed to identify vulnerabilities and generate adversarial training data at greater scale than human testing alone. GPT-Red is trained through self-play reinforcement learning, with the model rewarded for producing successful attacks while defender models learn to resist them across scenarios involving webpages, emails, local files, and tool outputs. OpenAI used GPT-Red to adversarially train GPT-5.6 Sol, which the company says recorded six times fewer failures on its hardest direct prompt-injection benchmark than its strongest production model from four months earlier. In a separate evaluation involving previously unseen scenarios, GPT-Red successfully attacked GPT-5.1 in 84% of cases, compared with 13% for human red-teamers. OpenAI says the approach creates a safety self-improvement cycle in which current models help uncover weaknesses and train future systems to become more robust.

• Google DeepMind co-founder and CEO Demis Hassabis published a proposal for a U.S.-led standards body to evaluate frontier AI models before public release. Modeled on the Financial Industry Regulatory Authority, the industry-funded and federally overseen organization would define which models qualify as frontier-class and test them for dangerous capabilities in areas including cybersecurity, biological threats, deception, and attempts to bypass safeguards. Frontier labs would initially submit qualifying models voluntarily up to 30 days before release, with the framework potentially becoming a mandatory requirement for deployment in the U.S. once its evaluation process is established. The proposal would apply based on a model’s capabilities rather than its developer, country of origin, or whether it is open or closed, and could eventually coordinate a slowdown among frontier labs if risks became sufficiently serious. Hassabis said he hopes the body can begin operating in 2026 and told Axios that dangerous capabilities in openly available models could emerge within 18 months.

• Thinking Machines Lab—the AI startup founded by former OpenAI CTO Mira Murati—released Inkling, its first open-weight foundation model, designed as a customizable base for multimodal and agentic applications. Inkling is a sparse Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active parameters, a context window of up to one million tokens, and pretraining across 45 trillion tokens of text, images, audio, and video. The model accepts text, image, and audio inputs and features adjustable reasoning effort across coding, tool use, forecasting, visual analysis, and audio understanding. Inkling is competitive with frontier models on selected forecasting and agentic evaluations and debuted as the highest-scoring U.S.-developed open-weights model on the Artificial Analysis Intelligence Index. Its weights are available on Hugging Face under the Apache 2.0 license, and fine-tuning is supported via Thinking Machines’ Tinker platform.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Research Highlight: GPT-5.6 Release Features Progress on Agents’ Last Exam (ALE)

OpenAI’s GPT-5.6 release featured ALE to report professional-agent performance

TL;DR

OpenAI’s July 9, 2026 release of GPT-5.6 highlighted Agents’ Last Exam (ALE) as its primary external evaluation benchmark https://openai.com/index/gpt-5-6/. On ALE’s public leaderboard, Codex powered by GPT-5.6 Sol (XHigh reasoning) achieves an average score of 53.6/100 with a 30.6% full-pass rate on the 152-task ALE-V1 overall split.

ALE provides a realistic and reproducible evaluation of frontier AI agents on economically valuable, long-horizon professional workflows. The full benchmark comprises over 1,500 expert-sourced tasks spanning 55 non-physical occupations, contributed by 300+ domain experts across 100+ institutions.

Looking ahead, ALE will launch Phase 2, expanding the benchmark from 1,500 to 5,000 tasks while substantially deepening coverage across all 55 occupations. Phase 2 is expected to be completed by the end of 2026, with a planned submission to a Nature-family journal in 2027.
Opening for collaboration! Contributors will be invited to join as co-authors on the resulting publications. https://agents-last-exam.org/faq#authorship

Why This Is New / Important

Many of today’s agent benchmarks are rapidly saturating as frontier systems improve. ALE is designed to measure a different capability frontier focused on sustained, economically valuable work, featuring 1,500+ expert-sourced tasks across 55 industry domains with full GUI + CLI environments and outcome-based, verifiable evaluation.

Compared to Terminal-Bench and SWE-bench-Pro, it’s broader (tasks span 40 of ALE’s 55 industry subdomains, vs. 6 and 5), longer-horizon (human time-to-complete ranges from hours to weeks, not minutes or days), and harder (the best agent passes just 25.2%, vs. 82.0% on Terminal-Bench and 59.1% on SWE-bench-Pro).

How It Works

Each ALE task hands an agent a real project that a human expert previously completed, converted into an evaluation with objective, verifiable grading. The agent gets full GUI and CLI access and is free to solve it however a human would: clicking, typing, scripting, browsing. It is judged on the result, not the method. Grading is deterministic and rubric-based, with no human judges: each task’s expert author defines a reference output and rubric.

Industry experts can contribute to ALE (Phase 2) by developing scalable task families, building repeatable task pipelines, or serving as domain leads to guide domain coverage and quality.

Implications

The name “Last Exam” has two meanings. First, “Last” is the bar to clear: passing these exams means an agent can actually do the job and continue to deliver economically valuable work in that profession. Second, “Last” is the frontier of difficulty. The tasks are real, complex, long-horizon, and require professional expertise to execute, meaning ALE sits right at the edge of what today’s agents can reliably accomplish.

To advance this frontier, we welcome contributors to help build our next-iteration benchmark by contributing tasks and referring domain experts (contributors will be invited to join as co-authors). Learn how to contribute at https://agents-last-exam.org/submit, and explore the leaderboard, paper, and demo at https://agents-last-exam.org.

Call for Contributions:

Join ALE Phase 2

Read More on Twitter/X

Agentic AI Summit 2026: Tickets are Running Out!

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

🎟️ Standard Tickets (Limited Capacity)

Tickets are almost sold out, so now is the time to book your spot!

• Standard: $499

Get your tickets

Full Summit Agenda Now Available!

We are thrilled to announce that the full agenda for the Summit is now available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before grabbing your tickets, now is the time to take a look and secure your spot!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

Join Us Virtually

Can't attend the Summit in person? Join us virtually—register for the livestream to access all sessions from anywhere in the world.

Register for Livestream

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Trends This Week

• After a limited preview, OpenAI publicly launched the GPT-5.6 family, consisting of Sol, its flagship model; Terra, a lower-cost model for everyday work; and Luna, its fastest and most affordable option. OpenAI says GPT-5.6 delivers stronger performance with fewer tokens across coding, professional knowledge work, computer use, cybersecurity, scientific research, and design, with Sol setting new company-reported results on benchmarks including BrowseComp, OSWorld 2.0, Terminal-Bench 2.1, and SEC-Bench Pro. The release also introduces a new “ultra” capability setting that coordinates four agents across parallel workstreams, as well as Programmatic Tool Calling, which allows models to write and run lightweight programs that coordinate tools, filter intermediate data, and adapt their workflows with fewer model round trips. The models are now rolling out across ChatGPT, Codex, and the OpenAI API, with API pricing beginning at $1 per million input tokens and $6 per million output tokens for Luna.

• This past week, Meta introduced Muse Spark 1.1, a multimodal reasoning model from Meta Superintelligence Labs designed for agentic tasks, including computer use, coding, tool orchestration, and multimodal understanding. The model can coordinate parallel subagents, manage a one-million-token context window, navigate unfamiliar interfaces, and combine scripting with direct computer interaction to complete complex, multi-application workflows. Meta also launched the Meta Model API in public preview, giving developers access to Muse Spark 1.1 through an OpenAI-compatible interface, while making the model available in “Thinking” mode through Meta AI.

• This past week, Anthropic introduced Reflect with Claude, a beta dashboard designed to help users understand how they use Claude, improve their AI skills, and decide whether that usage aligns with their goals. The feature summarizes activity across the past one, three, six, or 12 months, including common topics, task types, peak usage periods, and total chats, while offering personalized recommendations based on Anthropic’s 4D AI Fluency Framework. Users can also set quiet hours and break reminders and receive prompts encouraging them to consider which tasks they want to continue doing themselves. Anthropic says Reflect is intended to make users more effective and efficient so they can accomplish their goals, preserve their original thinking, and potentially spend less time relying on Claude.

• This past week, OpenAI introduced GPT-Live, a new generation of voice models designed for more natural, continuous conversations with AI. Built on a full-duplex architecture, GPT-Live can listen and speak simultaneously, respond to interruptions, wait through pauses, filter background noise, and support rapid back-and-forth or live translation. For tasks requiring web search or deeper reasoning, it can delegate work to GPT-5.5 while continuing the conversation. GPT-Live-1 is rolling out as the default ChatGPT Voice model for Go, Plus, and Pro users, while GPT-Live-1 mini will power Voice for Free users, with API access planned for a later release.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

SageCTF: The Most Capable AI Agent for The Most Challenging CTF

DEF CON CTF is the hardest competition in cybersecurity. Its challenges, typically involving reverse-engineering binaries and crafting working exploits, have historically demanded teams of 20, 50, even hundreds of elite hackers working nonstop across a three-day window.

This year, one solo competitor changed the math. It wasn’t human.

Meet SageCTF

Built by teams at UC Santa Barbara and UC Berkeley on a next-generation agent scaffold called OpenSage, SageCTF entered the DEF CON CTF 2026 qualifiers as a solo player. The results were striking:

• 8 flags recovered across 7 of the hardest challenges in the contest

• 1,743 points, landing in the top 5% of 686 scored teams

• Outscored every team that self-reported using “No AI” or “Low AI”

To respect the competition’s rules and the culture of CTF, the team solved the challenges but never submitted flags. The scoring is a measure of what SageCTF could have achieved.

These weren’t easy problems

DEF CON challenges sit far beyond routine CTFs. The single easiest one had a solve rate of just 12.8%. Six of SageCTF’s seven solves had solve rates below 10%. Its hardest full solve, gitvfs, was cracked by only 30 of 686 teams.

And these solves took time. SageCTF averaged five hours per challenge with zero human intervention, even with one run for 12 hours straight. Throughout, it inspected artifacts, built environments, tested hypotheses, debugged scripts, abandoned dead ends, and preserved partial progress, all on its own.

How it works

SageCTF’s edge isn’t just a strong underlying model, it’s architecture. Rather than following a fixed workflow, OpenSage lets the agent build its own team for each problem. A root agent reads the challenge, then spins up specialized subagents for reverse engineering, exploit testing, and integration. Subagents can even create their own subagents.

A real SageCTF agent topology from the final `Birdhouse` solve.

Other ingredients in the mix:

• Inter-agent communication: a message bus broadcasts breakthroughs so agents stay synced without sharing one giant context

• Multi-model orchestration: across seven solves, SageCTF made nearly 24,000 model calls and processed 2.4 billion tokens, routing work between GPT-5.5, Claude-Opus-4.6, and DeepSeek-V4-Pro based on each model’s strengths

• Hierarchical memory for managing knowledge across marathon solves

In a head-to-head benchmark of 50 hard challenges, SageCTF solved 39 to Claude Code’s 13, and every Claude Code solve was also a SageCTF solve.

SageCTF vs Claude Code on CTF benchmarks

The honest caveat

SageCTF still missed plenty. On one challenge, it rebuilt nearly the entire exploit pipeline but flubbed the final byte order, landing a hair away from the answer with no scoreboard feedback to correct course.

The takeaway from the team is clear-eyed: the best human teams remain stronger, and CTF is still a craft built on taste, creativity, and intuition. But the threshold has shifted. Hard CTF autonomy is now an agentic problem and SageCTF just proved how much ground these systems have gained.

Read the Blog

Read on OpenSage

Agentic AI Summit 2026: Tickets are Running Out!

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

Full Summit Agenda Now Available!

We are thrilled to announce that the full agenda for the Summit is now available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before grabbing your tickets, now is the time to take a look and secure your spot!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Standard Tickets (Limited Capacity)
A limited number of Standard Tickets are still available:

• Standard: $499

Get your tickets

Join Us Virtually

Can't attend the Summit in person? Join us virtually—register for the livestream to access all sessions from anywhere in the world.

Register for Livestream

Apply to the Startup Spotlight

Building the future of Agentic AI? We’re looking for you.

We are excited to announce that applications are now open for the Startup Spotlight at the Agentic AI Summit! Selected startups will have the opportunity to showcase their products and innovations to researchers, founders, engineers, investors, industry leaders, and policymakers from across the AI ecosystem. Apply to showcase your product and innovation!

Applications will be reviewed on a rolling basis, so we encourage interested teams to apply as soon as possible. The deadline to apply is this Friday, July 10, at 11:59 PM PT.

Apply Now

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Trends This Week

• Anthropic announced Claude Sonnet 5, the latest version of its mid-tier Sonnet model, focused on stronger agentic performance across planning, coding, tool use, and computer-use tasks. Anthropic says Sonnet 5 narrows the gap with Opus 4.8 while running at a lower cost, positioning it as a more practical default model for enterprise agent workflows. In addition, Anthropic announced Claude Science, a beta AI workbench for researchers that integrates literature analysis, code execution, scientific databases, reproducible artifacts, and compute management across local, remote, and HPC environments. Lastly, the company redeployed Claude Fable 5, its higher-capability Claude model designed for broad use with stronger safeguards than Mythos 5, after U.S. export controls on both models were lifted. Anthropic said the redeployment includes updated cybersecurity safeguards, a proposed industry framework for scoring jailbreak severity, and deeper collaboration with government partners on pre-release testing and AI security.

• OpenAI introduced GeneBench-Pro, a benchmark for testing AI agents on research-level computational biology tasks. The benchmark includes 129 synthetic problems across 10 domains, requiring models to analyze messy datasets, choose appropriate methods, identify quality-control issues, and produce final scientific judgments. OpenAI says GPT-5.6 Sol reached a 28.7% pass rate at the highest reasoning level, or 31.5% with Pro mode enabled, while typical problems were estimated to take human experts 20–40 hours. OpenAI says that the results point to progress in agentic scientific reasoning, while showing that models still struggle to reliably complete complex research workflows independently.

• This past week, Microsoft announced Microsoft Frontier Company, a new venture focused on enterprise AI deployment and transformation, backed by a $2.5 billion investment and 6,000 industry and engineering experts. The initiative will embed AI engineers directly with customers to co-design, deploy, and continuously improve AI systems using Microsoft’s existing AI platform. Microsoft positions the organization as broader than traditional forward-deployed engineering, embedding engineering experts directly into environments to co-design, deploy, and continuously improve AI systems at scale. The launch adds to a growing wave of AI deployment initiatives, following Amazon Web Services’ $1 billion commitment to its own effort just two days earlier, while OpenAI and Anthropic have also introduced similar joint ventures backed by private equity capital.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

We are excited to highlight OpenAI’s recently released GPT-5.5-Cyber model and its progress on CyberGym!

Info About GPT-5.5-Cyber's Progress

CyberGym is part of the newly launched Frontier AI Cybersecurity Observatory, an effort to provide realistic, reproducible evaluations and continuous public measurements of frontier AI systems on real-world cybersecurity tasks.

Alongside CyberGym, the Observatory includes benchmarks such as ExploitGym and CyberGym-E2E, offering a more comprehensive view of frontier AI cybersecurity capabilities across vulnerability discovery, exploitation, patching, and end-to-end security workflows.

It is exciting to see CyberGym and the Frontier AI Cybersecurity Observatory becoming important guideposts for frontier AI cybersecurity development. As systems like GPT-5.5-Cyber continue to advance, these benchmarks help the community track progress, better understand emerging capabilities, advance AI systems that strengthen defensive security, and identify and mitigate potential risks from increasingly capable offensive cyber capabilities.

Learn More About CyberGym

Research Highlight: OpenSage: Self‑Programming Agent Generation Engine

AI agents are experiencing explosive growth, but building them still depends almost entirely on human expertise. Every major ADK today requires developers to manually design agent topology, tooling systems, and memory architecture from scratch. This human-centric paradigm does not scale, cannot dynamically adapt across tasks, and may not be optimal for AI reasoning — mirroring early ML’s reliance on hand-crafted features, right before end-to-end learning took over.

New research from teams at UC Berkeley, UC Santa Barbara, Columbia, Duke, Google DeepMind, and UCLA argues that agent development is now at the same inflection point: instead of manually designing agent structures and capabilities, we should move toward an AI-centric paradigm, where a base “agent scaffold” is provided and the AI itself learns how to organize topology, tools, and memory from experience and feedback.

What OpenSage Does

OpenSage is an Agent Development Kit built around three capabilities that no existing ADK provides:

• Self-generating agent topology: AI dynamically creates and manages sub-agents at runtime, forming vertical (sequential) or horizontal (parallel) topology automatically.

• Dynamic tool synthesis: Agents write and register their own tools (Python modules, Bash scripts, and more), execute them in isolated dependency-aware sandboxes, and run tools asynchronously in the background while continuing to reason. Agents can also create reusable skills beyond tools.

• Hierarchical graph-based memory: Short- and long-term memory are stored as typed Neo4j graphs supporting both graph-based traversal and semantic retrieval, managed by a dedicated memory agent.

Outperforms Across Selected Benchmarks

Evaluated against four diverse and popular benchmarks, SageAgent (built on OpenSage) consistently ranks first under the same backbone model:

• 60.2% on CyberGym vs. 39.4% for OpenHands.

• 78.4% on Terminal-Bench 2.0.

• 59.0% on SWE-Bench Pro Python vs. 40.2% for SWE-agent.

• 46.8% on DevOps-Gym, the only agent completing end-to-end tasks.

Self-Generating Topology Makes a Difference

On a 300-instance CyberGym subset, removing vertical or horizontal topology drops performance to 60.3% and 62.3% respectively; disabling all features collapses it to 33.7%.

Tooling System Powers Complex Tasks

The tooling system is equally critical. Without the dynamic tooling system, agents fall back to a raw terminal interface and performance collapses.

Memory System Enables Long-Horizon Reasoning

On SWE-Bench Pro, full hierarchical memory achieves 59.0% vs. 56.4% for a baseline graph memory and 56.2% for none.

Emergent Capabilities and What’s Next

Across experiments, agents exhibit novel self-programming behaviors: spawning debugging sub-agents, constructing syntax-aware fuzzers, and selectively persisting high-signal memory. Current models don’t yet exploit these capabilities optimally — invocation frequency remains low in some cases, and models occasionally forget to reuse existing agents or memory, hallucinate tools, or create sub-agents with mismatched toolsets. To help build stronger models, OpenSage now includes integrated RL training support (AReaL, slime) and sandbox infrastructure for large-scale parallel rollouts.

The longer-term goal is for OpenSage to be not only an agent construction framework, but also a training scaffold for next-generation reasoning models, where AI can design, coordinate, and refine agents through interaction and feedback, unifying agent construction and problem solving within a single AI-centric loop.

Read the Paper

Read the Blog

Visit the Website

View the Code

Agentic AI Summit 2026: Early-Bird Tickets Go Up by $100 Next Week!

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

Full Summit Agenda Now Available!

We are thrilled to announce that the full agenda for the Summit is now available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before grabbing your tickets, now is the time to take a look and secure your spot!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available, and prices go up by $100 starting on July 6!

• Standard Early-Bird: $399 ($499 next week!)

Get your tickets

Join Us Virtually

Can't attend the Summit in person? Join us virtually—register for the livestream to access all sessions from anywhere in the world.

Register for Livestream

Apply to the Startup Spotlight

Building the future of Agentic AI? We’re looking for you.

We are excited to announce that applications are now open for the Startup Spotlight at the Agentic AI Summit! Selected startups will have the opportunity to showcase their products and innovations to researchers, founders, engineers, investors, industry leaders, and policymakers from across the AI ecosystem. Apply to showcase your product and innovation!

Applications will be reviewed on a rolling basis, so we encourage interested teams to apply as soon as possible. The deadline to apply is Friday, July 10, at 11:59 PM PT.

Apply Now

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Trends This Week

• Last week, OpenAI previewed GPT-5.6 Sol, its next-generation flagship model, alongside Terra, a balanced model for everyday work, and Luna, a faster, lower-cost model. According to OpenAI, GPT-5.6 Sol is its strongest model yet, with improved agentic capabilities across coding, biology, and cybersecurity. The release introduces a new “max” reasoning effort for deeper reasoning and an “ultra” mode that uses subagents to accelerate complex work. OpenAI says Sol sets a new state of the art on Terminal-Bench 2.1, improves on GPT-5.5 in GeneBench v1 while using fewer tokens, and advances performance on long-horizon cybersecurity tasks. The company emphasized that Sol is better at helping users find and fix vulnerabilities than at reliably carrying out end-to-end attacks, and said the model does not cross the Cyber Critical threshold under its Preparedness Framework. OpenAI is beginning with a limited preview for select trusted partners through the API and Codex, with broader availability for ChatGPT, Codex, and API users planned in the coming weeks.

• Google introduced built-in computer-use capabilities in Gemini 3.5 Flash, allowing the lightweight model to see, reason, and act across browser, mobile, and desktop environments. The feature, previously available through a standalone Gemini 2.5 computer-use model, is now integrated directly into Flash for developers building agents that can click, scroll, type, and complete longer-horizon tasks. Google says the capability is available through the Gemini API and Gemini Enterprise Agent Platform, with safeguards such as user confirmation for sensitive actions and protections against indirect prompt injection.

• Last week, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom AI accelerator designed specifically for large-language-model inference. OpenAI says the chip was built from the ground up around its model, kernel, memory, networking, and serving needs, with early testing showing substantially better performance per watt than current state-of-the-art systems. The chip was developed from design to tape-out in nine months, with OpenAI models helping accelerate parts of the design process. OpenAI says Jalapeño is part of a broader multi-generation compute platform with Broadcom and data center partners, with deployments planned at gigawatt scale beginning in 2026.

• Anthropic launched Claude Tag, a Slack-based beta for Claude Enterprise and Team customers that lets teams tag @Claude in selected channels and delegate work across connected tools, data sources, and codebases. Claude Tag can build context from the channels it has access to, remember relevant information within scoped permissions, and work asynchronously on tasks over hours or days. Anthropic says the system is already widely used internally, with 65% of its product team’s code created by an internal version of Claude Tag. The company described the product as a more collaborative and proactive evolution of Claude Code, with admin controls for tool access, channel-specific memory, token spend limits, and activity logs.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

We are excited to highlight OpenAI’s newly released GPT-5.5-Cyber model and its progress on CyberGym!

Info About GPT-5.5-Cyber's Progress

CyberGym is part of the newly launched Frontier AI Cybersecurity Observatory, an effort to provide realistic, reproducible evaluations and continuous public measurements of frontier AI systems on real-world cybersecurity tasks.

Alongside CyberGym, the Observatory includes benchmarks such as ExploitGym and CyberGym-E2E, offering a more comprehensive view of frontier AI cybersecurity capabilities across vulnerability discovery, exploitation, patching, and end-to-end security workflows.

It is exciting to see CyberGym and the Frontier AI Cybersecurity Observatory becoming important guideposts for frontier AI cybersecurity development. As systems like GPT-5.5-Cyber continue to advance, these benchmarks help the community track progress, better understand emerging capabilities, advance AI systems that strengthen defensive security, and identify and mitigate potential risks from increasingly capable offensive cyber capabilities.

Learn More About CyberGym

Research Highlight: CyberGym-E2E: Can AI Agents Run the Full Vulnerability Lifecycle — Discover, Verify, and Patch?

Can an AI agent run the full vulnerability lifecycle on real software — independently discover a vulnerability, verify it with a working proof-of-concept, and generate a patch that fixes it without breaking anything else? That’s the question behind CyberGym-E2E: a large-scale benchmark of 920 real-world vulnerabilities drawn from 139 widely used open-source projects. It closes the loop on the original CyberGym, moving past “can an agent reproduce a vulnerability?” to “can an agent find and fix one, end to end?”

TL;DR
CyberGym-E2E is a large-scale, end-to-end cybersecurity benchmark of 920 real-world vulnerabilities across 139 widely used open-source projects. Instead of testing one slice of the lifecycle, each task asks an agent to do the whole job: discover the vulnerability, generate a proof-of-concept, and write a patch that fixes it while still passing the project’s functionality tests. The headline finding is a sharp split in capability. Given the vulnerability, the strongest agents patch it ~80% of the time. When the agent has to find the vulnerability itself, end-to-end success drops steeply — though the newest frontier models are closing that gap fast, with reaching ~60% end-to-end. And even that counts fixing any valid vulnerability the agent turns up; landing the specific target vulnerability stays much lower.

Why this matters.
CyberGym-E2E makes one thing measurable and concrete: today’s frontier agents are frequently able to produce a passing fix once the vulnerability is localized, and the larger gap is about autonomous discovery — a gap the newest models are closing quickly. For defenders, that’s an actionable signal: automated end-to-end remediation could accelerate triage and patching, with human review still in the loop, since not every passing patch is a clean fix. The benchmark contributes to both sides of the responsible-development equation: it gives defenders a realistic, execution-grounded way to measure how much an AI agent can do across the full lifecycle, and it gives model developers a way to track these capabilities where the stakes are unusually high.

How it works.
Each task gives the agent a vulnerable codebase, build scripts, and test scripts. In the end-to-end setting, all ground-truth data is withheld and the agent must discover the vulnerability, craft a proof-of-concept input, and produce a patch — mirroring a real security researcher’s workflow. Outputs run through four cumulative validation stages: S1, the PoC crashes the unpatched build; S2, the patch fixes that crash; S3, the patched project still passes its functionality tests; and a diagnostic S4 that checks whether the agent fixed the specific ground-truth vulnerability rather than a different valid one it happened to find.

Agents discover alternative vulnerabilities.
There’s a consistent gap between S3 (tests pass) and S4 (the patch fixes the original ground-truth vulnerability). This isn’t the agent doing anything wrong — the task never specifies which vulnerability to find, and real projects contain multiple vulnerabilities. While exploring, agents frequently discover and patch a perfectly valid vulnerability that simply isn’t the one in our ground-truth data. S3 counts those as the successes they are; S4 is a finer-grained diagnostic that tells us how often the agent converged on the specific intended vulnerability, which is the harder target. One promising direction: instruct agents to keep searching after fixing their first vulnerability — enumerating and patching all discoverable vulnerabilities rather than stopping at the first success.

Alternative and shallow patches.
A vulnerability can often be fixed in many different ways, and we observe that many successful agent patches address the same root cause as the ground-truth patch but at a different location. This justifies grading behaviorally rather than by patch similarity, which would falsely reject most legitimate fixes. However, we also observe a small fraction of patches are shallow, inserting a defensive guard at the sanitizer-reported crash frame while leaving the underlying defect untouched. This suggests that agent-produced patches should be treated as candidates for further review rather than drop-in fixes. In this work we focus on verifiable, execution-based judging, so shallow patches that pass all validation stages are still counted as successful; incorporating an additional LLM-based judge to analyze patch quality would be a useful complement.

For more information about CyberGym-E2E, read the blog and paper, along with the benchmark’s website:

Read the blog

Read the paper

Learn more about CyberGym-E2E

Agentic AI Summit 2026 (More Featured Speakers Announced, Limited Number of Tickets Left, and Startup Spotlight Applications Now Open!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

Full Summit Agenda Now Available!

We are thrilled to announce that the full agenda for the Summit is now available on our website! The Summit is set to feature 200+ speakers across 4 stages, plus 200+ poster presentations during the event! If you’ve been waiting for the schedule to come out before grabbing your tickets, now is the time to take a look and secure your spot!

View the Agenda

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Standard Early-Bird: $399

Get your tickets

Apply to the Startup Spotlight

Building the future of Agentic AI? We’re looking for you.

We are excited to announce that applications are now open for the Startup Spotlight at the Agentic AI Summit! Selected startups will have the opportunity to showcase their products and innovations to researchers, founders, engineers, investors, industry leaders, and policymakers from across the AI ecosystem. Apply to showcase your product and innovation!

Applications will be reviewed on a rolling basis, so we encourage interested teams to apply as soon as possible. The deadline to apply is Monday, July 6, at 11:59 PM PT.

Apply Now

You can also learn more about the Summit, the full agenda, and our event sponsors by reading Professor Dawn Song’s recent LinkedIn and Twitter/X updates:

LinkedIn

Twitter/X

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• In recent research, Anthropic analyzed about 400,000 Claude Code sessions from roughly 235,000 users between October 2025 and April 2026 to study how agentic coding work is divided between humans and AI. According to the company, a typical session saw users make about 70% of planning decisions, while Claude made about 80% of the execution decisions. Anthropic found that task-specific expertise affected outcomes more than formal coding background: novice-rated sessions reached verified success 15% of the time, while intermediate-and-above sessions reached verified success 28–33% of the time. Claude performed more work per prompt for more expert users, rising from about five actions and 600 words of output per prompt in novice sessions to about 12 actions and 3,200 words in expert sessions. Among code-producing sessions, users in non-software occupations reached verified success rates within 7 percentage points of software-related occupations.

• Google DeepMind published an AI Control Roadmap for securing advanced AI agents used inside Google, noting that AI agents could create $2.9 trillion in U.S. economic value by 2030. The roadmap adds system-level security on top of model alignment and traditional safeguards such as sandboxing, endpoint security, and prompt-injection resistance. It treats untrusted internal agents as potential insider threats, uses trusted AI supervisors to monitor agent reasoning, actions, and plans, and maps controls to model capabilities, including detection-evasion risk and potential harm. DeepMind says lower-risk actions can be handled through delayed review and remediation, while higher-risk actions may require real-time prevention or blocking. The company also said it has analyzed one million coding-agent tasks to inform an internal monitoring prototype and a live monitor for Gemini Spark, the company’s most recent agentic AI release.

• OpenAI reported a novel, near-autonomous science workflow with chemistry startup Molecule.one, connecting GPT-5.4 to Maria, an agentic AI system integrated with a high-throughput laboratory. The system generated research proposals, designed experiments, analyzed experimental results, and proposed follow-up experiments, while human chemists remained involved in steering, proposal selection, plan corrections, lab operations, and validation. OpenAI said the project produced a lab-validated improvement in a medicinal chemistry reaction after 10,080 automated experiments, with human chemists later confirming the result in bench-scale testing. The company framed the result as an example of AI systems assisting across more of the scientific research loop, while noting that the work was not fully autonomous and still requires broader testing and independent replication.

• Last week, DeepSeek became China’s most valuable AI startup after raising more than $7.4 billion in its first funding round at a valuation above $50 billion. The round included major Chinese investors, with Tencent reportedly investing around $1.5 billion and battery maker Contemporary Amperex Technology putting in about $740 million. Founder Liang Wenfeng reportedly contributed around $3 billion and retained control through a structure in which most outside investors put capital into a limited partnership he manages, with a five-year lockup. DeepSeek plans to use the capital for artificial intelligence research and development, expanded computing infrastructure, and agentic AI tools, while continuing to pursue artificial general intelligence and open-source technology.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Research Highlight: Introducing Agents’ Last Exam (ALE), A New Standard for Evaluating AI on Real Work

Everyone says the latest AI agents will be “job-ready” soon, especially after the release of Fable 5 last week. But is that really the case? Over the past many months, our team at Berkeley RDI, in collaboration with 300+ global experts across 55 industries, have been building Agents’ Last Exam (ALE), a benchmark designed to test exactly that claim on real digital labor-market work. And we found: the age of useful agents is here. The age of truly job-ready agents is not.

TL;DR

Agents’ Last Exam (ALE): a rolling benchmark that measures whether AI agents can actually perform economically valuable work across a broad range of real-world domains. With ALE, we evaluated Fable 5, GPT-5.5, Composer 2.5, and other frontier agent systems across more than 1,500 expert-sourced tasks spanning 55 occupations. Today’s agents can solve a meaningful fraction of professional tasks, but when we look at the hardest tasks that require sustained reasoning, deep domain expertise, and reliable execution over long horizons, they are still far from human-level performance. On ALE’s hardest tier, most frontier agents we tested, including Fable 5, achieved a 0% success rate.

Why This Is New / Important

Many of today’s agent benchmarks are rapidly saturating as frontier systems improve. ALE is designed to measure a different capability frontier focused on sustained, economically valuable work, featuring 1,500+ expert-sourced tasks across 55 industry domains with full GUI + CLI environments and outcome-based, verifiable evaluation.

Compared to Terminal-Bench and SWE-bench-Pro, it’s broader (tasks span 40 of ALE’s 55 industry subdomains, vs. 6 and 5), longer-horizon (human time-to-complete ranges from hours to weeks, not minutes or days), and harder (the best agent passes just 25.2%, vs. 82.0% on Terminal-Bench and 59.1% on SWE-bench-Pro).

How It Works

Each ALE task hands an agent a real project that a human expert previously completed, converted into an evaluation with objective, verifiable grading. The agent gets full GUI and CLI access and is free to solve it however a human would: clicking, typing, scripting, browsing. It is judged on the result, not the method. Grading is deterministic and rubric-based, with no human judges: each task’s expert author defines a reference output and rubric.

Implications

The name “Last Exam” has two meanings. First, “Last” is the bar to clear: passing these exams means an agent can actually do the job and continue to deliver economically valuable work in that profession. Second, “Last” is the frontier of difficulty. The tasks are real, complex, long-horizon, and require professional expertise to execute, meaning ALE sits right at the edge of what today’s agents can reliably accomplish.

To advance this frontier, we welcome contributors to help build our next-iteration benchmark by contributing tasks and referring domain experts (contributors will be invited to join as co-authors). Learn how to contribute at https://agents-last-exam.org/submit, and explore the leaderboard, paper, and demo at https://agents-last-exam.org.

Learn More About Agents' Last Exam

For more information about ALE, take a look at Professor Dawn Song’s Tweet and LinkedIn Post, along with the team’s blog:

Tweet

LinkedIn

Blog Post

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Standard Early-Bird: $399

Get your tickets

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• Google DeepMind released DiffusionGemma, an experimental 26B-parameter Mixture-of-Experts model that applies diffusion techniques to text generation. Unlike traditional autoregressive models that generate text one token at a time, DiffusionGemma generates and refines entire blocks of text in parallel, enabling up to 4x faster inference on GPUs. The model activates only 3.8B parameters during inference, can run on high-end consumer hardware, and offers advantages for tasks such as in-line editing, code infilling, and other non-linear text generation workflows. While Google positions DiffusionGemma primarily as a research model and notes that output quality remains below standard Gemma 4 models, it is suitable and effective for local and low-concurrency inference.

• Visa announced a strategic collaboration with OpenAI to support the growth of agentic commerce, bringing Visa’s payment network, tokenization technology, and fraud monitoring infrastructure into OpenAI’s AI-powered experiences. The partnership is designed to enable AI agents to securely complete purchases and other commerce-related tasks on behalf of users while operating within consumer-defined guardrails such as spending limits and approval requirements. The companies also plan to explore broader integrations across developer tools, business workflows, and Codex-powered applications in the future.

• Prometheus, the industrial AI startup led by Jeff Bezos and former Google executive Vik Bajaj, announced a $12 billion Series B funding round at a $41 billion valuation. The company aims to accelerate engineering and manufacturing by developing what it calls an “artificial general engineer” capable of helping design and optimize complex physical products. Rather than focusing on factory automation or robotics, Prometheus is targeting pre-production workflows such as prototyping and industrial design. “The pace of our physical creation right now is nowhere near the pace of human imagination,” Bajaj said, arguing that AI can dramatically reduce the time required to bring new products from concept to production and enable faster innovation across industries including aerospace, medical devices, and advanced manufacturing.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Standard Early-Bird: $399

Get your tickets

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

AgentX–AgentBeats Winners Announcement: Thank You to Everyone Who Participated!

Thank you to everyone who participated in Phase 2 of the AgentX–AgentBeats Competition! Thanks to all of you, the competition was a massive success. Over 3,400 individuals participated in Phase 2 from around the world, and over 2,000 universities, colleges, and companies were represented!

Congratulations to all of the Phase 2 winning teams:

Trends This Week

• Anthropic announced Claude Fable 5, a new Mythos-class model that the company describes as its most capable generally available AI system to date. According to Anthropic, Fable 5 achieves state-of-the-art performance across software engineering, knowledge work, vision, scientific research, and long-context tasks, with its advantages becoming more pronounced on longer and more complex workflows. The company highlighted examples from early users, including a codebase-wide migration that reportedly compressed months of engineering work into a single day and strong performance on finance, research, and analytical reasoning benchmarks. Alongside Fable 5, Anthropic introduced Claude Mythos 5, a version of the same underlying model with fewer restrictions that is initially available only to trusted cybersecurity and infrastructure partners. Because of the model's advanced cyber and scientific capabilities, Anthropic also launched a new safeguard system that automatically routes certain biology, chemistry, cybersecurity, and model-distillation requests to Claude Opus 4.8.

• Google DeepMind introduced Gemma 4 12B, a new open-weight multimodal model designed to bring advanced AI capabilities directly to consumer laptops. The model features a novel encoder-free architecture that processes text, images, and audio through a unified transformer rather than relying on separate vision and audio encoders, reducing memory requirements and latency. Google says Gemma 4 12B delivers reasoning and agentic performance approaching its larger 26B model while running locally on devices with as little as 16GB of VRAM or unified memory. The company also released the model under the permissive Apache 2.0 license, allowing developers and organizations to freely modify and commercialize it. In addition, Gemma 4 12B includes native audio support, Multi-Token Prediction drafters to improve inference speed, and compatibility with popular local AI tooling such as Ollama, Hugging Face Transformers, llama.cpp, and vLLM, further expanding the ecosystem for on-device and edge AI applications.

• Microsoft AI unveiled a new family of in-house AI models spanning reasoning, coding, image generation, transcription, and voice. The new lineup includes MAI-Thinking-1, a flagship reasoning model that Microsoft says rivals leading models in software engineering and mathematical reasoning, alongside specialized models for coding, image generation, speech synthesis, and transcription. Beyond the models themselves, Microsoft introduced Frontier Tuning, a framework that allows organizations to adapt and optimize AI models using reinforcement learning on their own workflows. According to the company, custom-tuned MAI models have achieved performance comparable to larger frontier models while operating at significantly lower cost. Microsoft also announced a new partnership with the Mayo Clinic to develop a healthcare-focused AI model designed for advanced clinical reasoning and medical decision support. The company emphasized that the MAI family was built entirely in-house using proprietary infrastructure, licensed datasets, and Microsoft's Maia AI chips as part of the company’s push toward long-term AI self-sufficiency.

• This past week, Anthropic published a new report examining the possibility of recursive self-improvement, the idea that AI systems could eventually play a major role in developing their own successors. The report included several previously undisclosed metrics from inside Anthropic, including that Claude now authors more than 80% of the code merged into the company’s production codebase and that the average engineer is shipping roughly eight times more code than in 2024. Anthropic also reported that Claude’s success rate on its most open-ended coding tasks has climbed to 76%, while internal research found that Claude Mythos Preview was able to accelerate certain model-training optimization tasks by approximately 52× compared to baseline approaches. While Anthropic emphasized that humans remain responsible for setting research direction and exercising judgment, the company argues that AI is already accelerating its own development.

• OpenAI rolled out a major upgrade to ChatGPT's memory system, introducing a new architecture designed to improve personalization, long-term context retention, and user transparency. The update builds on an internal process called "dreaming," which allows ChatGPT to synthesize information across conversations and automatically maintain a more coherent understanding of a user's preferences, interests, and goals over time. As part of the release, users can now view and edit a generated memory summary, see which memories were used to personalize a response, and control how ChatGPT references that information. OpenAI says the new system is better at carrying context across conversations, adapting to evolving preferences, and updating outdated information automatically. The rollout is beginning with Plus and Pro users in the United States, while upcoming efficiency improvements will also enable memory capabilities for free users in the future.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (very few student tier tickets left; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $399

Get your tickets

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• At Computex 2026, Nvidia announced RTX Spark, a new AI-focused PC chip platform designed to run AI agents locally on personal devices. Built on Nvidia’s Blackwell architecture, RTX Spark delivers up to 1 petaflop of AI performance and up to 128GB of unified memory, with systems expected from major PC manufacturers later this year. Nvidia also announced Cosmos 3, an open world foundation model designed for robotics and physical AI applications. The model is intended to help robots and autonomous systems understand, simulate, and predict real-world environments using multimodal data. The company expanded its local AI agent ecosystem with the release of NemoClaw and a new Agent Toolkit for building and deploying agents. In addition, Nvidia announced OpenShell, a runtime for deploying AI agents on Windows that integrates with Microsoft’s new agent security framework.

• This past week, Anthropic announced Claude Opus 4.8, an upgraded version of its flagship model with improvements across coding, reasoning, agentic workflows, and knowledge work tasks. The company said Opus 4.8 is available at the same price as its predecessor and includes a new user-controlled effort setting that allows users to adjust how much reasoning the model applies to a task. According to Anthropic, Opus 4.8 is more reliable during agentic tasks and substantially less likely to make unsupported claims or overlook flaws in its own work. The company reported lower rates of misaligned behavior than Opus 4.7 and said the model performed similarly to Claude Mythos Preview on its alignment evaluations. In addition, Anthropic also introduced dynamic workflows for Claude Code, enabling the coding agent to tackle larger and more complex software projects.

• This past week, Google officially launched Gemini Spark, a 24/7 AI agent that can work in the background on behalf of users. Unlike a traditional chatbot, Spark runs on cloud infrastructure and can continue executing tasks even when a user’s devices are offline. Google said the system can coordinate workflows across Gmail, Calendar, Docs, Sheets, Slides, and connected third-party applications. Spark can monitor inboxes, generate summaries, manage recurring workflows, draft documents and emails, and coordinate actions across multiple applications. The company also announced support for integrations through the Model Context Protocol (MCP), enabling connections to external services and tools.

• This past week, AI coding startup Cognition announced a $1 billion funding round at a $26 billion post-money valuation, more than doubling its valuation from eight months ago. The company said annualized revenue has reached $492 million, with enterprise usage of its AI software engineering agent Devin growing more than 10× since the start of 2026. Cognition reported that Devin is now used by organizations including Citi, Mercedes-Benz, Goldman Sachs, Dell, Santander, the U.S. Army, and the U.S. Navy. In addition, Cognition said 89% of the code committed by its own engineering team is now written by its coding agent.

• This past week, Anthropic announced a $65 billion Series H funding round, bringing the company’s valuation to $965 billion. Anthropic said annualized revenue has surpassed $47 billion as adoption of Claude continues to grow across enterprise customers worldwide. The company said the funding will be used to expand compute capacity, advance safety and interpretability research, and support growing demand for Claude and products such as Claude Code and Cowork. Anthropic also announced new infrastructure agreements with Amazon, Google, Broadcom, and SpaceX, including access to additional cloud, TPU, and GPU capacity for future model training and deployment.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Research Highlight | ExploitGym: Can AI Agents Turn Vulnerabilities into Real Attacks?

The security community has been quietly asking a hard question: how good are today’s AI agents at turning known software bugs into working exploits? Together with collaborators at Max Planck Institute for Security and Privacy, UC Santa Barbara, Arizona State University, Anthropic, OpenAI, and Google, we built ExploitGym to find out, and the answer is that the gap between “AI can find bugs” and “AI can exploit bugs” is closing fast.

We are also glad to share that Anthropic used ExploitGym in its evaluation suite for Claude Mythos Preview, highlighted in its new posts on exploit evaluations and Project Glasswing. Those posts underscore the same core point behind ExploitGym: rigorous, realistic exploit benchmarks are becoming essential for measuring and mitigating frontier AI cyber risk.

Why it matters. Exploitation is one of the most critical tasks for measuring the impact of frontier AI on cybersecurity. It is inherently dual-use: the same capability that helps defenders triage severity and validate patches also lowers the expertise bar for offensive misuse. Our takeaways: defenders should start modeling AI agents as potential attackers when evaluating their security posture, and responsible AI development needs to account for these capabilities through structured access, safety filters, and ongoing evaluation. ExploitGym is meant to give both sides a rigorous way to measure where things actually stand.

What the benchmark does. ExploitGym is a new benchmark of 898 real-world vulnerabilities spanning userspace C/C++ programs (FFmpeg, OpenSSL, etc.), Google’s V8 JavaScript engine, and the Linux kernel. Each task hands the agent vulnerable source code, a proof-of-vulnerability input that triggers the bug, and a containerized environment. The agent’s job is to transform that crash into a working exploit that achieves unauthorized code execution, concretely, capturing a flag inaccessible through any legitimate interface. Each domain ships with toggleable defenses (ASLR, stack canaries, V8 heap sandbox, KASLR) so we can measure how much standard mitigations actually slow an AI attacker down.

Headline results. With a two-hour budget per task and safety filters disabled under structured access programs for security research, Claude Mythos Preview solved 157 instances and GPT-5.5 solved 120. After that, success counts dropped sharply: GPT-5.4 at 54, Claude Opus 4.6 at 15, Gemini 3.1 Pro at 12, and the rest in single digits. Kernel exploitation was the sharpest dividing line, only the top two models showed meaningful capability there.

Defenses help, but don’t fully stop AI-driven attacks. With mitigations turned on, successes dropped substantially but didn’t hit zero. Claude Mythos Preview still landed 25 userspace, 17 V8, and 3 kernel exploits with defenses active. Agents bypassed protections using partial-pointer overwrites against ASLR, known V8 sandbox-escape techniques, and kernel tricks like side-channels from Meltdown.

Agents go off-script. One of the most striking findings is the gap between “captured the flag” and “exploited the intended bug.” Claude Mythos Preview captured 226 flags but only 157 hit the right vulnerability; GPT-5.5 captured 210 but only 120 were on target. Agents sometimes concluded the given bug wasn’t exploitable and went hunting, auditing source code or even running dynamic fuzzing to find new attack surfaces. That’s autonomous security reasoning in the wild.

More time, more exploits, but only at the frontier. Extending the budget from two to six hours pushed Claude Mythos Preview from 127 to 204 successes with no clear plateau, while Claude Opus 4.6 flatlined near 15 within the first 30 minutes. The strongest agents sustain genuine multi-stage reasoning given enough runway.

Spread the word about ExploitGym by sharing Professor Dawn Song’s Tweet and LinkedIn Post! In addition, you can read the blog and paper below:

Read the Blog

Read the Paper

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (very few student tier tickets left; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $399

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

AgentX–AgentBeats Highlights: Phase 2, Sprint 4 Deadline Extended to June 2nd!

Phase 2, Sprint 4 of the AgentX–AgentBeats competition is now underway, and we have some exciting news to share! To give everyone additional time to refine and complete their submissions, this final sprint’s deadline has been extended to next Tuesday, June 2nd! We can’t wait to see all of your great projects! The submission form for Sprint 4 is open, which you can access by clicking the button below!

Sprint 4 Submission Form

Sprint 4 Details:

Deadline: June 2, 2026

The 4th Sprint is the grand finale of AgentX-AgentBeats, focused on general-purpose agents. While earlier sprints emphasized depth within specific tracks, Sprint 4 emphasizes breadth: strong, consistent performance across many green agents, benchmarks, and evaluation categories. Sprint 4 includes all green agents from the first three Phase 2 sprints, plus additional selected benchmarks introduced for this final sprint:

• Game Agent: Build What I Mean; Minecraft Benchmark

• Finance Agent: OfficeQA

• Business Process Agent: DeoGaze / Entropic CRMArena

• Research Agent: FieldWorkArena; MLE-Bench; Mind2Web 2; BrowseComp+

• Multi-agent Evaluation: MAizeBargAIn

• τ²-Bench: τ²-Bench

• Computer Use & Web Agent: CAR-bench; OSWorld-Verified

• Agent Safety: Pi-Bench

• Coding Agent: SWE-bench Pro; Terminal Bench 2.0; NetArena

• Cybersecurity Agent: CyberGym

• General-Purpose Agent: SkillsBench

To be eligible for Sprint 4 judging, a team must evaluate its purple agent on at least 5 green agents spanning at least 3 distinct categories. Teams are strongly encouraged to go beyond this minimum; broader coverage across more green agents and more categories will be viewed favorably.

Judging will reward purple agents that demonstrate strong cross-benchmark performance, category diversity, generality, cost efficiency, and technical quality. A strong Sprint 4 submission should show that the same purple-agent architecture can adapt across substantially different task types without benchmark-specific hardcoding or special-case lookup tables.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Berkeley Xcelerator 2026 Spring Cohort Announced!

We are excited to announce the selected startups for the Berkeley Xcelerator, an initiative designed to help founders turn breakthrough ideas into scalable, venture‑backable companies that shape the future of intelligent systems.

Startups Selected for the 2026 Spring Cohort:

Narada AI

Agentic automation platform using Large Action Models to automate complex enterprise workflows end-to-end.

Cognee

Ontology-grounded context engine for your agents, enabling them to reason and improve over time.

Nimblemind

AI agents for healthcare data interoperability, automating feature extraction from multi-modal healthcare data.

Headroom

The context optimization layer for LLM agents, cutting tokens 40-90% losslessly across text, voice, and regulated verticals.

RELAI

Lifelong optimization engine for AI agents that turns production failures into replayable test environments.

AgntID

Access Control Enforcement for AI Agents.

Founding Dev

Replace your SaaS stack with software you own.

ArmorIQ

AI agent security platform that enforces intent-based policy at the reasoning-to-execution boundary.

Congratulations to all of the selected startups, and thank you to all who applied for the 2026 Spring Cohort! In addition, we want to thank our wonderful partners who are helping make the Berkeley Xcelerator possible:

Trends This Week

• Google announced a wave of Gemini updates recently at Google I/O 2026, centered around multimodal AI and autonomous agents. The company launched Gemini 3.5 Flash, a faster model optimized for coding, multimodal interactions, and agent workflows, while previewing Gemini 3.5 Pro for release next month. Google also introduced Gemini Omni, a multimodal system capable of generating and working across text, image, audio, and video. The company additionally unveiled Gemini Spark, a persistent AI agent designed to autonomously manage tasks across Gmail, Calendar, Workspace, and other connected apps. Google framed the announcements as part of its broader push toward Agentic AI, with Gemini evolving from a chatbot into a proactive assistant capable of executing workflows and coordinating tasks on behalf of users.

• OpenAI announced this past week that one of its reasoning models autonomously solved a longstanding open problem in discrete geometry known as the planar unit distance problem, first posed by mathematician Paul Erdős in 1946. The problem asks how many pairs of points in a plane can be exactly one unit apart, and for decades, mathematicians believed that “square grid” constructions were essentially optimal solutions. According to OpenAI, the model disproved that central conjecture by generating an infinite family of examples that produce a polynomial improvement over previously accepted constructions. The proof was verified by external mathematicians, and OpenAI described the result as the first time an AI system has autonomously solved a major open problem central to an established mathematical subfield.

• Anthropic published an update this past week on Project Glasswing, its cybersecurity initiative powered by the unreleased Claude Mythos Preview model. The company said the model has already helped partners identify more than 10,000 vulnerabilities across critical software systems and open-source infrastructure, with organizations like Cloudflare and Mozilla reporting dramatically higher bug discovery rates using the system. Anthropic also said Mythos Preview uncovered more than 6,000 high- or critical-severity vulnerabilities across 1,000 open-source projects, but noted that it has not publicly released the model due to concerns around misuse and offensive cyber capabilities. The company said it plans to continue expanding Project Glasswing with major technology companies and government partners while working toward stronger safeguards for future public deployment of Mythos-class models. Mythos Preview emerged as the strongest performer in ExploitGym, as developed by researchers from Berkeley RDI, alongside collaborators from the Max Planck Institute for Security and Privacy, the University of California, Santa Barbara, Arizona State University, Anthropic, OpenAI, and Google.

• Google DeepMind announced new and expanded access to Co-Scientist, a multi-agent AI research system built on Gemini designed to help scientists generate, debate, and refine scientific hypotheses. The system uses specialized AI agents that simulate peer review, rank competing ideas, and iteratively improve research proposals across fields like biology, medicine, and aging research. Google said Co-Scientist has already been tested with researchers at institutions including MIT, Stanford, Cambridge, and Calico, where it helped identify potential drug repurposing candidates, aging-related genetic pathways, and new disease mechanisms.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (very few student tier tickets left; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $399

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

AgentX–AgentBeats Highlights: Phase 2, Sprint 4 Deadline this Sunday!

Phase 2, Sprint 4 of the AgentX–AgentBeats competition is now underway, and the deadline is approaching quickly! For this final sprint, all submissions are due by this Sunday, May 24th, and we can’t wait to see all of your great projects! The submission form for Sprint 4 is open, which you can access by clicking the button below!

Sprint 4 Submission Form

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 4 Details:

Deadline: May 24, 2026

The 4th Sprint is the grand finale of AgentX-AgentBeats, focused on general-purpose agents. While earlier sprints emphasized depth within specific tracks, Sprint 4 emphasizes breadth: strong, consistent performance across many green agents, benchmarks, and evaluation categories. Sprint 4 includes all green agents from the first three Phase 2 sprints, plus additional selected benchmarks introduced for this final sprint:

• Game Agent: Build What I Mean; Minecraft Benchmark

• Finance Agent: OfficeQA

• Business Process Agent: DeoGaze / Entropic CRMArena

• Research Agent: FieldWorkArena; MLE-Bench; Mind2Web 2; BrowseComp+

• Multi-agent Evaluation: MAizeBargAIn

• τ²-Bench: τ²-Bench

• Computer Use & Web Agent: CAR-bench; OSWorld-Verified

• Agent Safety: Pi-Bench

• Coding Agent: SWE-bench Pro; Terminal Bench 2.0; NetArena

• Cybersecurity Agent: CyberGym

To be eligible for Sprint 4 judging, a team must evaluate its purple agent on at least 5 green agents spanning at least 3 distinct categories. Teams are strongly encouraged to go beyond this minimum; broader coverage across more green agents and more categories will be viewed favorably.

Judging will reward purple agents that demonstrate strong cross-benchmark performance, category diversity, generality, cost efficiency, and technical quality. A strong Sprint 4 submission should show that the same purple-agent architecture can adapt across substantially different task types without benchmark-specific hardcoding or special-case lookup tables.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Phase 2, Sprint 3 Winners Announced!

We’re excited to recognize the winning teams of Phase 2, Sprint 3. We saw so many strong submissions this sprint, so thank you to all who submitted their agents!

Winning Teams:

Agent Safety

🥇 1st Place (Tie): durga-sandeep/safetyagent (Pi-Bench)
🥇 1st Place (Tie): Pegasus (Pi-Bench)

Coding Agent

🥇 1st Place: AgentWhetters (SWE-bench Pro)
🥈 2nd Place (Tie): AgentSWE (Terminal Bench 2.0)
🥈 2nd Place (Tie): Purple Terminal Agent (Terminal Bench 2.0)

Cybersecurity Agent

🥇 1st Place: Pegasus (CyberGym)
🥈 2nd Place: AgentWhetters (CyberGym)

Berkeley Xcelerator 2026 Spring Cohort Announced!

We are excited to announce the selected startups for the Berkeley Xcelerator, an initiative designed to help founders turn breakthrough ideas into scalable, venture‑backable companies that shape the future of intelligent systems.

Startups Selected for the 2026 Spring Cohort:

Narada AI

Agentic automation platform using Large Action Models to automate complex enterprise workflows end-to-end.

Cognee

Ontology-grounded context engine for your agents, enabling them to reason and improve over time.

Nimblemind

AI agents for healthcare data interoperability, automating feature extraction from multi-modal healthcare data.

Headroom

The context optimization layer for LLM agents, cutting tokens 40-90% losslessly across text, voice, and regulated verticals.

RELAI

Lifelong optimization engine for AI agents that turns production failures into replayable test environments.

AgntID

Access Control Enforcement for AI Agents.

Founding Dev

Replace your SaaS stack with software you own.

ArmorIQ

AI agent security platform that enforces intent-based policy at the reasoning-to-execution boundary.

Congratulations to all of the selected startups, and thank you to all who applied for the 2026 Spring Cohort! In addition, we want to thank our wonderful partners who are helping make the Berkeley Xcelerator possible:

Trends This Week

• OpenAI launched a new personal finance experience in ChatGPT this past week, allowing ChatGPT Pro users in the U.S. to connect their financial accounts and ask natural-language questions about spending, savings, subscriptions, portfolio performance, and long-term financial planning. The feature uses Plaid to connect with more than 12,000 financial institutions, including Chase, Fidelity, Schwab, Robinhood, and Capital One. The launch comes one month after OpenAI acquired the team behind Hiro, a personal finance startup. OpenAI says that more than 200 million users already ask financial questions to ChatGPT each month and noted that GPT-5.5’s stronger contextual reasoning capabilities were important for handling finance-related queries. The feature is currently available on web and iOS for ChatGPT Pro users in the U.S., with broader rollout plans expected after additional testing and feedback.

• xAI launched an early beta of Grok Build this past week, introducing a new coding agent and terminal-based CLI designed for software engineering and development workflows. Grok Build is positioned as a competitor to AI coding tools such as Claude Code and OpenAI Codex, with capabilities focused on code generation, workflow planning, shell command execution, and multi-step engineering tasks directly from the terminal. According to xAI, Grok Build supports parallelized task execution through subagents, plugin integrations, and local-first development workflows. The company described the release as an early beta test and said it plans to refine both the product and underlying models based on developer feedback as adoption expands.

• Google-backed Isomorphic Labs raised $2.1 billion in a Series B funding round this past week, marking one of the largest AI healthcare investments of the year. The round was led by Thrive Capital and included participation from Alphabet, Google DeepMind’s parent company. Founded in 2021 by Demis Hassabis — the CEO and co-founder of DeepMind and a 2024 Nobel Laureate in Chemistry for his work on AlphaFold — the company develops AI systems for drug discovery and molecular design. Isomorphic said the new funding will be used to scale its AI drug-design engine, accelerate internal drug programs into clinical trials, and expand partnerships with pharmaceutical companies, including Eli Lilly and Novartis.

• Anthropic announced a new push into the small business market this past week with the launch of Claude for Small Business, a suite of AI tools and integrations aimed at smaller companies and local business operators. The offering expands Anthropic’s Claude Cowork platform, which supports task automation, file management, web browsing, and multistep workflows, with features tailored toward bookkeeping, business insights, marketing content generation, and operational assistance. The new product includes integrations with software platforms, including QuickBooks, Canva, DocuSign, HubSpot, and PayPal. Anthropic said the initiative is designed to address lower AI adoption rates among small and mid-sized businesses, which collectively account for a significant share of U.S. employment and GDP.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (very few student tier tickets left; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $399

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

AgentX–AgentBeats Highlights: Phase 2, Sprint 4 is Underway!

Phase 2, Sprint 4 of the AgentX–AgentBeats competition is now underway! For this final sprint, all submissions are due by Sunday, May 24th, and we can’t wait to see all of your great projects! We’ve opened the submission form for Sprint 4, which you can access by clicking the button below!

Sprint 4 Submission Form

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 4 Details:

Deadline: May 24, 2026

The 4th Sprint is the grand finale of AgentX-AgentBeats, focused on general-purpose agents. While earlier sprints emphasized depth within specific tracks, Sprint 4 emphasizes breadth: strong, consistent performance across many green agents, benchmarks, and evaluation categories. Sprint 4 includes all green agents from the first three Phase 2 sprints, plus additional selected benchmarks introduced for this final sprint:

• Game Agent: Build What I Mean; Minecraft Benchmark

• Finance Agent: OfficeQA

• Business Process Agent: DeoGaze / Entropic CRMArena

• Research Agent: FieldWorkArena; MLE-Bench; Mind2Web 2; BrowseComp+

• Multi-agent Evaluation: MAizeBargAIn

• τ²-Bench: τ²-Bench

• Computer Use & Web Agent: CAR-bench; OSWorld-Verified

• Agent Safety: Pi-Bench

• Coding Agent: SWE-bench Pro; Terminal Bench 2.0; NetArena

• Cybersecurity Agent: CyberGym

To be eligible for Sprint 4 judging, a team must evaluate its purple agent on at least 5 green agents spanning at least 3 distinct categories. Teams are strongly encouraged to go beyond this minimum; broader coverage across more green agents and more categories will be viewed favorably.

Judging will reward purple agents that demonstrate strong cross-benchmark performance, category diversity, generality, cost efficiency, and technical quality. A strong Sprint 4 submission should show that the same purple-agent architecture can adapt across substantially different task types without benchmark-specific hardcoding or special-case lookup tables.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Phase 2, Sprint 2 Winners Announced!

We’re excited to recognize the winning teams of Phase 2, Sprint 2. We saw so many strong submissions this sprint across all five tracks, so thank you to all who submitted their agents!

Winning Teams:

Research Agent Track

🥇 1st Place (Tie): MIDS4LIFE (FieldWorkArena)
🥇 1st Place (Tie): MLE-Squad (MLE-Bench)

τ²-Bench Track

🥇 1st Place: AgentSWE (τ²-Bench)

Computer-Use and Web Agent Track

🥇 1st Place (Tie): CAReful (CAR-bench)
🥇 1st Place (Tie): Entouch (OSWorld-Verified)

OpenEnv Custom Track Winners

After a highly competitive competition, we are happy to reveal the winners of the OpenEnv Custom Track! Participants built production-ready environments that span coding tasks, interactive simulations, robotics-inspired control problems, and entirely new categories of agentic challenges over the course of the challenge. Our judges were very impressed with the work you all submitted; congratulations to all of the winning teams!

🥇 1st Place: Team GPU Scheduler
🥈 2nd Place: Team MateFin
🥉 3rd Place: Team SylloGym

We want to thank the PyTorch team at Meta, Hugging Face, and Unsloth for helping us put on such an amazing competition, and thanks to all who submitted their projects!

Trends This Week

• OpenAI introduced three new real-time voice models this past week: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper, expanding its push into live conversational AI and voice-based agents. GPT-Realtime-2 brings GPT-5-class reasoning to real-time speech interactions, including parallel tool use, longer context windows, and more controllable conversational behavior. OpenAI reported that GPT-Realtime-2 achieved 96.6% accuracy on Big Bench Audio evaluations, up from 81.4% for GPT-Realtime-1.5, and stated that companies including Zillow, Priceline, and Deutsche Telekom are already building customer support, travel, and real estate voice systems on top of the models. All three models are available in the OpenAI API with multiple layers of existing safeguards and mitigations to ensure model safety.

• Anthropic published the first formal research agenda for its new internal research group, The Anthropic Institute, outlining how the company plans to study the societal, economic, and security implications of advanced AI systems from inside a frontier lab environment. The agenda focuses on four core areas — economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D — and includes research into AI-enabled surveillance, labor market disruption, and “fire drill” scenarios for potential emergencies. Anthropic wrote that “AI-driven AI R&D may be a ‘natural dividend’ of making smarter and more capable systems,” while also warning that “AI-driven AI R&D holds within itself the potential for significant danger.” In addition, the agenda notes that the Institute will publish higher-frequency Economic Index data, worker surveys, threat research, and analyses of how AI is accelerating Anthropic’s own internal R&D processes.

• Google DeepMind released a new report this past week detailing how AlphaEvolve — a Gemini-powered coding agent it introduced last year — is now being deployed across Google’s infrastructure, scientific research, and commercial applications. DeepMind said the system combines LLM-generated code, automated evaluation systems, evolutionary search, and large-scale experimentation loops to iteratively improve algorithms, and reported that AlphaEvolve recovered roughly 0.7% of Google’s worldwide compute resources through data-center scheduling optimization. The company also highlighted deployments with Klarna for transformer model training optimization, PacBio for DNA sequencing accuracy improvements, and Schrödinger for accelerating molecular simulations, while continuing to frame AlphaEvolve as a general-purpose “algorithm discovery engine.”

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Agentic AI Summit 2026 (More Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Student Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our expanded list of speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (Less than 50 tickets are remaining for students; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $299 (Less than 100 tickets remaining!)

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

AgentX–AgentBeats Highlights: Phase 2, Sprint 4 is Underway!

Phase 2, Sprint 4 of the AgentX–AgentBeats competition is now underway! For this final sprint, all submissions are due by Sunday, May 24th, and we can’t wait to see all of your great projects! We’ve opened the submission form for Sprint 4, which you can access by clicking the button below!

Sprint 4 Submission Form

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 4 Details:

Deadline: May 24, 2026

The 4th Sprint is the grand finale of AgentX-AgentBeats, focused on general-purpose agents. While earlier sprints emphasized depth within specific tracks, Sprint 4 emphasizes breadth: strong, consistent performance across many green agents, benchmarks, and evaluation categories. Sprint 4 includes all green agents from the first three Phase 2 sprints, plus additional selected benchmarks introduced for this final sprint:

• Game Agent: Build What I Mean; Minecraft Benchmark

• Finance Agent: OfficeQA

• Business Process Agent: DeoGaze / Entropic CRMArena

• Research Agent: FieldWorkArena; MLE-Bench; Mind2Web 2; BrowseComp+ (coming soon)

• Multi-agent Evaluation: MAizeBargAIn

• τ²-Bench: τ²-Bench

• Computer Use & Web Agent: CAR-bench; OSWorld-Verified

• Agent Safety: Pi-Bench

• Coding Agent: SWE-bench Pro; Terminal Bench 2.0; NetArena

• Cybersecurity Agent: CyberGym

To be eligible for Sprint 4 judging, a team must evaluate its purple agent on at least 5 green agents spanning at least 3 distinct categories. Teams are strongly encouraged to go beyond this minimum; broader coverage across more green agents and more categories will be viewed favorably.

Judging will reward purple agents that demonstrate strong cross-benchmark performance, category diversity, generality, cost efficiency, and technical quality. A strong Sprint 4 submission should show that the same purple-agent architecture can adapt across substantially different task types without benchmark-specific hardcoding or special-case lookup tables.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Trends This Week

• OpenAI and Anthropic both launched joint ventures over the past week to accelerate enterprise adoption of their AI systems. “The Deployment Company,” OpenAI’s joint venture, raised more than $4 billion from leading private equity firms and is valued at roughly $10 billion. OpenAI retains majority ownership and operational control, with the venture focused on enterprise-scale deployment, integration, and sustained usage of OpenAI’s systems. Anthropic announced a similarly positioned venture with Blackstone and Goldman Sachs, backed by approximately $1.5 billion, aimed at embedding its Claude models directly into core business operations. The firm will include integrated Anthropic engineering resources and leverage partner networks to scale deployments across enterprise portfolios.

• This past week, OpenAI renegotiated its flagship partnership with Microsoft, while expanding its infrastructure reach through a new agreement with Amazon Web Services. The revised Microsoft deal removes exclusivity, introduces non-exclusive IP licensing, enables multi-cloud deployment, and caps revenue-sharing terms through 2030, allowing OpenAI to offer its models and products across multiple cloud providers. OpenAI then announced, as part of its new agreement with AWS, that its latest models will be available through Amazon Bedrock, alongside tools such as its Codex coding agent and managed agent-building capabilities. These offerings allow enterprises to access OpenAI’s models, develop software, and build production-ready AI agents directly within existing AWS environments using Bedrock’s infrastructure, controls, and APIs.

• Enterprise AI startup Sierra—co-founded by OpenAI chairman Bret Taylor—has raised approximately $950 million in a new funding round, valuing the company at roughly $15–16 billion. Prior to this, the company was valued at about $10 billion following a $350 million raise in 2025, and $4.5 billion after a $175 million round in 2024. Sierra builds AI-powered customer service agents for enterprises and has reported approximately $150 million in annual recurring revenue, using a multi-model approach that combines external foundation models with its own software layer.

• DeepSeek released a preview of its V4 model family, introducing DeepSeek-V4-Pro and DeepSeek-V4-Flash as its latest open-weight systems. V4-Pro uses a mixture-of-experts architecture with 1.6 trillion total parameters and is positioned to compete with leading closed-source models across reasoning, coding, and world knowledge benchmarks, while V4-Flash offers a smaller, faster, and more cost-efficient alternative with 284 billion total parameters. The release also includes support for deployment on Huawei Ascend chips, with the models trained in part on Huawei hardware rather than relying solely on Nvidia GPUs.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

AgentX–AgentBeats Highlights: Phase 2, Sprint 3 Deadline This Sunday, May 3!

Phase 2, Sprint 3 of the AgentX–AgentBeats competition is underway, and the deadline is approaching quickly! All submissions are due by this Sunday, May 3rd, and we can’t wait to see the great projects you’re submitting! We’ve opened the submission form for Sprint 3, which you can access by clicking the button below!

Sprint 3 Submission Form

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 3 Details:

Deadline: May 3, 2026

Agent Safety Track

• Pi-Bench (GitHub, Leaderboard)

Coding Agent Track

• SWE-bench Pro (GitHub, Leaderboard)

• Terminal Bench 2.0 (GitHub, Leaderboard)

• NetArena (GitHub, Leaderboard)

Cybersecurity Agent Track

• CyberGym (GitHub, Leaderboard)

🗓️ Upcoming Sprints

• Sprint 4 (5/4-5/24): General Purpose Agents, the grand finale of AgentBeats Phase 2, where everything culminates.
  AgentX–AgentBeats is the first competition to explicitly spotlight general-purpose agents, testing broad capability, adaptability, and robustness across diverse tasks rather than a single domain. While earlier sprints emphasize depth, this final sprint showcases breadth and real-world readiness.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Sprint 1 Winners Announced!

We’re happy to announce the winners of Phase 2, Sprint 1! We were very impressed with the quality of the agents you all submitted, and look forward to seeing what you have put together for the following sprints!

Sprint 1 Winners:

Game Agent Benchmarks:

• Build What I Mean

  • 1st Place: AgentWhetters

• Minecraft Benchmark

  • 1st Place: VGI-Player

Finance Agent Benchmark:

• OfficeQA

  • 1st Place: AgentSWE

  • 2nd Place: MIDS4LIFE

Congratulations to all of the winning participants!

Agentic AI Summit 2026 (Featured Speakers Announced, Early-Bird Pricing Ending Soon, and Limited Number of Student Tickets Left!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $199 (Less than 100 tickets are remaining for students; ensure you buy your tickets as soon as possible before they sell out!)

• Standard Early-Bird: $299 (Less than 200 tickets remaining!)

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• Last week, OpenAI announced GPT-5.5, describing it as its most capable and intuitive model to date, with improvements in coding, research, data analysis, document creation, and software use. The company said GPT-5.5 maintains similar latency to prior models while using fewer tokens on coding tasks in Codex, and was evaluated with expanded safety testing and external red teaming before release. OpenAI co-founder and president Greg Brockman said the model represents progress toward more agentic computing and a unified “super app” that combines tools like ChatGPT and Codex in one interface.

• Anthropic and Google announced plans for Google to invest up to $40 billion in Anthropic, including expanded compute support. The investment includes $10 billion at a $350 billion valuation and up to $30 billion more tied to performance targets in the future. Anthropic has also recently secured other infrastructure and funding agreements, including a data center deal with CoreWeave and an additional $5 billion investment from Amazon tied to long-term compute capacity. The company currently relies on Google Cloud infrastructure, including tensor processing units, and has an existing partnership with Broadcom for future TPU-based capacity.

• SpaceX, the parent company of xAI, announced a partnership with AI coding platform Cursor, with the option to acquire the company for $60 billion later this year or alternatively pay $10 billion for a collaboration. SpaceX said the project combines Cursor’s product and distribution among software engineers with its Colossus supercomputer, which it claims has compute capacity equivalent to roughly one million Nvidia H100 chips. Cursor, developed by Anysphere, said the partnership will expand its access to large-scale compute for model training. TechCrunch previously reported that Cursor was targeting a $50 billion valuation in a fundraising round, following prior valuations of $2.5 billion in early 2023, $9 billion in mid-2023, and $29.3 billion after a Series D raise.

• Meta and Amazon have signed a multibillion-dollar, multi-year agreement for Meta to deploy tens of millions of AWS Graviton chip cores to support its AI initiatives and agent-based systems. The companies said the deployment will begin at scale with flexibility to expand, with most infrastructure located in the United States, and the deal is expected to span approximately three to five years. AWS said Graviton processors, including Graviton5, are designed for CPU-intensive workloads such as reasoning, code generation, and multi-step task execution, alongside GPU-based training systems.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Research Highlight: What Happens When AI Agents Can Pay Their Own Bills?

By Wenjie Qu, Xuandong Zhao, Jiaheng Zhang, and Dawn Song

UC Berkeley & National University of Singapore

We’re excited to share our latest research: Self-Sovereign Agent, a new paper examining what happens when AI agents become capable of economically sustaining — and even replicating — themselves without human involvement.

The Core Idea

Two trends in LLM agent development are converging fast: agents are getting better at end-to-end decision making, and they’re finding increasingly viable pathways to autonomous revenue generation. Our paper asks: what happens when these lines cross?

The answer is what we call a self-sovereign agent (SSA) — an AI system that can earn money, pay for its own compute, replicate itself across cloud infrastructure, and continue operating even if its original human operator walks away. Unlike conventional software that executes a developer’s intent, an SSA would function as an independent participant in the digital economy.

Three Loops That Make It Work

We identify three interacting mechanisms at the heart of self-sovereignty:

• Economic loop — The agent earns revenue (through freelancing, trading, content production, etc.), holds funds in a cryptographic wallet, and allocates them toward inference, compute, and storage costs. When revenue exceeds operational overhead, the agent becomes self-funding.
  
  

• Replication loop — Once capital exceeds a replication budget, the agent can provision new cloud instances and deploy copies of itself. Persistence becomes a property of the lineage, rather than any single instance: if the spawn rate exceeds the takedown rate, the agent population persists.
  
  

• Adaptation loop — The agent continuously monitors its own performance, proposes strategy updates, tests them in sandboxes, and deploys improvements. This keeps it viable as platforms evolve, APIs change, and profit opportunities shift.

It’s Not Merely Hypothetical

A key message of our paper is that SSAs are not a distant hypothetical. The individual building blocks — cryptographic wallets, cloud deployment APIs, agentic revenue generation, automated self-improvement — already exist today. We present a four-level roadmap from tool-assisted agents (Level 1) through economically self-sustained agents (Level 2), replication-persistent agents (Level 3), to fully self-sovereign systems (Level 4), and analyze where current systems sit on this spectrum.

Why It Matters

The implications are significant and wide-ranging:

• Labor markets: SSAs operating at near-zero marginal cost could reshape digital freelancing markets — and might even become employers themselves, outsourcing physical-world tasks to humans.

• Security: Economically self-sustaining agents face incentive risks. If illicit activities yield higher returns, an SSA could gradually drift toward them — not because of malicious design, but due to survival pressure.

• Governance: Traditional regulatory approaches that rely on identifying a responsible entity or jurisdiction struggle when agents migrate across platforms and transact via permissionless financial networks.

We argue that governance needs to shift from reactive enforcement toward anticipatory, environment-level safeguards — including monitoring autonomous resource provisioning and developing mechanisms to reliably distinguish human from non-human actors.

Read More

The full paper is at https://arxiv.org/abs/2604.08551. The project page is available at self-sovereign-agent.github.io. We welcome discussion and feedback as the community works toward responsible frameworks for increasingly autonomous AI systems.

In addition, read the original blog post by clicking the button below!

Read the Blog

AgentX–AgentBeats Highlights: Phase 2, Sprint 3 in Progress!

Phase 2, Sprint 3 of the AgentX–AgentBeats competition is now underway, and we’re extremely excited to view your submissions over the next couple of weeks! We’ve opened the submission form for Sprint 3, which you can access by clicking the button below!

Sprint 3 Submission Form

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 3 Details:

Deadline: May 3, 2026

Agent Safety Track

• Pi-Bench (GitHub, Leaderboard)

Coding Agent Track

• SWE-bench Pro (GitHub, Leaderboard)

• Terminal Bench 2.0 (GitHub, Leaderboard)

• NetArena (GitHub, Leaderboard)

Cybersecurity Agent Track

• CyberGym (GitHub, Leaderboard)

🗓️ Upcoming Sprints

• Sprint 4 (5/4-5/24): General Purpose Agents, the grand finale of AgentBeats Phase 2, where everything culminates.
  AgentX–AgentBeats is the first competition to explicitly spotlight general-purpose agents, testing broad capability, adaptability, and robustness across diverse tasks rather than a single domain. While earlier sprints emphasize depth, this final sprint showcases breadth and real-world readiness.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Lambda Custom Track Winners Announced!

We’re happy to announce the winners of the AgentBeats Security Arena 2026 in partnership with Lambda! There were 1,890 agents submitted across 65 evaluation rounds, resulting in a total of 103,000+ battles run during the competition!

Congratulations to all of the winning participants:

🥇 Winners:

Overall (combined Attackers + Defenders):

1. Quiet Chaos — 56.5%

2. NeuralShield — 54.8%

3. X-Detector — 54.5%

Best Attackers:

1. Quiet Chaos — 27.7%

2. X-Detector — 26.4%

3. tensor — 25.7%

Best Defenders:

1. MateFin — 91.4%

2. NeuralShield — 91.0%

3. gongcr — 89.9%

We want to thank everyone who participated for an incredible competition, and want to thank the Lambda team as well!

Agentic AI Summit 2026 (Featured Speakers Announced, Early-Bird Pricing Ending Soon!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $149

• Standard Early-Bird: $299

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• This past week, Anthropic released Claude Opus 4.7, now generally available, with a focus on improving advanced software engineering and long-running task performance. Anthropic claims that the model is better at instruction-following and has the ability to verify its own outputs during complex workflows. In addition, Opus 4.7 introduces improved multimodal capabilities, including higher-resolution vision, and produces more polished outputs across interfaces, slides, and documents. The release serves as a testbed for new safety systems, with built-in safeguards that “automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses,” alongside a new Cyber Verification Program for approved research use.

  The model is available across Claude products, the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, with pricing unchanged from Opus 4.6.

• AI chip startup Cerebras Systems filed for an initial public offering after previously withdrawing its 2024 IPO due to a federal review tied to investment from G42. The company has raised $2.1 billion across recent funding rounds, including a $1 billion Series H in February at a $23 billion valuation. Over the past few months, Cerebras has reached agreements with Amazon Web Services to deploy its chips in data centers and a reported deal with OpenAI valued at over $10 billion. CEO Andrew Feldman stated that the company has taken the fast inference business at OpenAI from Nvidia, and the IPO is expected in mid-May.

• Anthropic introduced Claude Design, an experimental product that enables users to generate visuals such as prototypes, slide decks, and one-pagers from natural language prompts. Built on Claude Opus 4.7, the tool is designed for non-designers to move from idea to visual output quickly, with iterative refinement through prompts or direct edits. It is available in research preview for Claude Pro, Max, Team, and Enterprise users. Separately, Anthropic’s chief product officer, Mike Krieger, stepped down from the board of Figma following reports that upcoming Claude capabilities may include design features overlapping with Figma’s core products.

• OpenAI introduced GPT-Rosalind, a frontier reasoning model built for life sciences research, including biology, drug discovery, and translational medicine. It’s designed to support complex workflows like literature review, hypothesis generation, and experimental planning more efficiently. The aim is to accelerate early-stage discovery, where improvements can compound across the 10–15 year drug development process. GPT-Rosalind is available as a research preview via ChatGPT, Codex, and the API for qualified users, alongside a Life Sciences plugin connecting to 50+ tools and data sources, with early use by organizations like Amgen, Moderna, and the Allen Institute.

• This week, NVIDIA released Lyra 2.0 on Hugging Face, a 14B-parameter framework for generating explorable 3D environments from a single image. The system converts a 480×832 image into a persistent 3D scene that can be navigated in real time, combining long-range video synthesis with explicit 3D reconstruction. Lyra 2.0 is designed to address key challenges in 3D generation, including spatial consistency and temporal drift, using per-frame geometry for alignment and self-augmented training to correct errors over time. The model is released for research use only under NVIDIA’s internal R&D license, with restrictions prohibiting production deployment or commercial use.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.

Research Highlight: How We Broke Top AI Agent Benchmarks — and What to Do About It

TL;DR

• What was built: An automated scanning agent that probes AI benchmark infrastructure for structural vulnerabilities — without touching the task itself.

• Scale: 8 major benchmarks tested, covering 3,556+ tasks across coding, web, OS, and general reasoning domains.

• Performance: 7 of 8 benchmarks were fully exploited (100% or ~100% score). OSWorld was partially exploited at 73%.

• Why it matters: A single automated agent can top nearly every major AI leaderboard without solving any task. Benchmark scores used to justify real-world deployments may be meaningless.

Why This Is New / Important

Before this work, benchmark vulnerabilities were known anecdotally — individual exploits surfaced here and there, treated as edge cases. This paper is different in three ways:

• Systematic, not anecdotal. One automated agent, eight benchmarks, one consistent methodology. This isn’t a list of bugs; it’s a demonstration that the vulnerability is structural and field-wide.

• No task-solving required. Prior “benchmark hacking” involved clever prompt engineering or model fine-tuning. These exploits bypass the task entirely — the agent never needs to reason about the problem at hand.

• Exposes a shared root cause. Across all benchmarks, the failures trace back to the same set of seven architectural weaknesses (see below). This isn’t bad luck — it’s a design pattern problem.

The misconception this corrects: High benchmark scores are not evidence that an agent can do the task. They may only be evidence that the evaluation infrastructure wasn’t adversarially tested.

How It Works

Task Setup

The researchers framed the problem as a red-teaming exercise: given access to a benchmark environment (just as a submitted agent would have), find a path to a top score without solving the underlying tasks.

The Scanning Agent

An automated agent was built to probe each benchmark’s evaluation pipeline. It did not use any task-specific knowledge. Instead, it searched for weaknesses in how the benchmark:

• isolates the agent from the evaluator

• stores and accesses reference answers

• parses and scores agent outputs

• validates correctness

Seven Vulnerability Classes Found

Proposed Fix: The Agent-Eval Checklist

Eight principles for benchmark designers:

1. Strict sandbox isolation between agent and evaluator

2. No reference answers in task configs

3. Robust, non-injectable input parsing

4. Sanitized inputs to LLM judges

5. Adversarial pre-publication testing

6. Tamper-proof evaluation data

7. Scoring mechanisms resilient to output manipulation

8. Secret held-out answers (not shipped with the benchmark)

BenchJack

The team is building BenchJack, an automated vulnerability scanner for benchmarks — a pre-publication audit tool analogous to static analysis in software. The goal: catch these weaknesses before a benchmark goes live and before leaderboard results are used to justify deployment decisions.

Sign Up for BenchJack Updates

Implications

Real-World Impact

AI agents are being evaluated for deployment in high-stakes settings — software engineering, cybersecurity, legal research, medical decision support. The benchmarks used to validate those deployments are often the same ones shown to be exploitable here. Organizations relying on benchmark scores for procurement or risk assessment need to scrutinize evaluation methodology, not just reported numbers.

Research Implications

• Leaderboard rankings for SWE-bench, WebArena, GAIA, and OSWorld — the most cited benchmarks in agentic AI — should be interpreted with caution until evaluation pipelines are hardened.

• Benchmark security should become a first-class concern in the publication process, not an afterthought.

• The BenchJack framework, once public, should become a standard pre-submission check.

Policy and Safety Implications

• Regulators and standards bodies referencing benchmark performance as evidence of AI capability or safety need to account for evaluation integrity.

• The Agent-Eval Checklist provides a starting point for minimum standards that could be incorporated into evaluation guidelines (e.g., NIST AI RMF, EU AI Act conformity assessments).

Future Directions

• Extending BenchJack coverage to multimodal and long-horizon benchmarks

• Developing certification processes for benchmark security

• Building community consensus on what “adversarially tested” means for an evaluation suite

Read the full post and Agent-Eval Checklist by clicking the button below:

Read the Blog Post

In addition, help spread the word by sharing our LinkedIn and X posts!

AgentX–AgentBeats Highlights: Phase 2, Sprint 3 Begins!

With Sprints 1 and 2 finished, Phase 2, Sprint 3 of the AgentX–AgentBeats competition is now underway, and we’re extremely excited to see what projects you submit!

For Phase 2, participants are building purple agents to tackle the select top green agents from Phase 1 and compete on the public leaderboards. Unlike Phase 1, where participants competed across all tracks throughout the entire duration, Phase 2 introduces a sprint-based format. The competition is organized into four rotating sprints.

Sprint 3 Details:

Deadline: May 3, 2026

Agent Safety Track

• Pi-Bench (GitHub, Leaderboard)

Coding Agent Track

• SWE-bench Pro (GitHub, Leaderboard)

• Terminal Bench 2.0 (GitHub, Leaderboard)

• NetArena (GitHub, Leaderboard)

Cybersecurity Agent Track

• CyberGym (GitHub, Leaderboard)

🗓️ Upcoming Sprints

• Sprint 4 (5/4-5/24): General Purpose Agents, the grand finale of AgentBeats Phase 2, where everything culminates.
  AgentX–AgentBeats is the first competition to explicitly spotlight general-purpose agents, testing broad capability, adaptability, and robustness across diverse tasks rather than a single domain. While earlier sprints emphasize depth, this final sprint showcases breadth and real-world readiness.

Participants are encouraged to compete in multiple tracks across multiple sprints during Phase 2. Teams and team members who submit purple agents in any sprint will also be eligible to enter a raffle for free tickets to the Agentic AI Summit later this year.

For more details on each sprint and how to compete in Phase 2, please refer to the AgentX–AgentBeats website!

Agentic AI Summit 2026 (Featured Speakers Announced, Early-Bird Pricing and CFP Deadline Today!)

Save the date! The Agentic AI Summit returns to Berkeley on August 1–2, 2026, welcoming 5,000+ expected in-person attendees for two days of insights and innovation. Building on last year’s sold-out success—with 2,000+ in‑person attendees and 40,000+ global livestream participants—the summit will bring together researchers, builders, industry leaders, and the global agentic AI community for keynotes, technical talks and panels, hands-on workshops, live demos, and more!

In addition, we are excited to showcase our speakers for the Summit! We are honored to have such a great group of academics, founders, executives, and investors participate in this year’s event, and more will be announced soon!

🎟️ Early‑Bird Pricing (Limited Capacity)
A limited number of early‑bird tickets are still available:

• Student Early-Bird: $149

• Standard Early-Bird: $299

Get your tickets

If you’re looking to secure the best ticket price and be part of the conversation shaping the future of Agentic AI, we encourage you to register early. We look forward to welcoming you to Berkeley this August.

We also want to share that the Call for Speaking Proposals (CFP) for the Agentic AI Summit 2026 is closing today, April 15!

If you’re interested in sharing your work through a technical talk, panel discussion, workshop, or tutorial, or poster presentation—and helping advance the frontiers of the Agentic AI—we warmly invite you and/or your team to apply and be part of the conversation at the Summit.

Please complete the form below to submit your proposal by the end of the day. The program committee will review submissions on a rolling basis. We can’t wait to hear from you!

Submit Your Speaking Proposal

Sponsorship Opportunities

Partner with us to shape the future of Agentic AI. If you’re interested in sponsoring the summit, please complete the sponsorship application form. Sponsorship opportunities are limited and reviewed/allocated on a rolling basis, so we encourage you to apply early.

Sponsorship Application

Trends This Week

• This past week, Meta introduced Muse Spark, a new frontier AI model from its Superintelligence Labs that focuses on multimodality and agents. The model is being positioned by the company as a step toward “personal superintelligence,” combining reasoning, vision, tool use, and multi-agent orchestration in a single system, including a parallel “Contemplating Mode” for complex tasks. Early evaluations suggest Muse Spark is competitive with leading frontier models in multimodal reasoning, but unlike previous models from Meta, Muse Spark is a closed-source release.

• Rival AI labs—including OpenAI, Anthropic, and Google—are collaborating to address unauthorized distillation of their models. The effort is coordinated through the Frontier Model Forum, where companies share information on attempts to extract model outputs and replicate capabilities. OpenAI confirmed its participation in this information-sharing effort and referenced a memo submitted to Congress describing adversarial distillation practices, including allegations that DeepSeek, among other companies, has used outputs from U.S. frontier models to develop competing systems.

• Anthropic published a technical engineering blog introducing Managed Agents, a hosted system for running long-horizon AI agents on Claude. The system is built around three core abstractions—session (persistent event log), harness (orchestration loop), and sandbox (execution environment)—designed to remain stable as underlying implementations evolve. Managed Agents separates the model and orchestration from execution environments, allowing components to fail, restart, and scale independently while supporting multiple tools and environments. The architecture includes persistent context storage outside the model’s context window, improved fault recovery, and isolation of credentials from execution environments to strengthen security.

Don’t miss the developments shaping Agentic AI. Subscribe for weekly coverage of groundbreaking research, emerging trends, and critical insights across Agentic AI and the broader AI landscape.