Over the weekend I updated my Second Brain setup. If you member I wrote about building local LLMWiki on Andrej Karpathy’s idea of an LLM maintaining its own wiki. I completed the restructuring today, and the changes are worth a short note.

The old setup produced an individual page for every ingested story from a single ingest batch. The single flat wiki folder is now split into four: concepts, entities, articles, and early signals. Every distinct story gets its own page, deduplicated by its source URL, and weak single-source mentions are kept in a separate “signals” folder so they capture and documented appropriately.

The bigger change is underneath the content. I added a gated tooling layer: the scripts that regenerate links and rebuild the index, audits the wiki hygiene on a regular basis, are now versioned, and changes to any of them only ships if it passes a full lint and validation gate and produces identical output on a repeat run. I built this after reading Google’s WikiSkill paper, which argues that an agent system holds up better long-term when raw activity, accumulated knowledge, and executable skills are kept as three separate layers, with the skill layer only updated once it clears a real test. That distinction mapped directly onto a bug I had already hit: a maintenance script rebuilt from OpenClaw memory had silently reintroduced an old error. Gating changes behind a lint pass and an idempotence check is the direct fix.

I also wrote an explicit operating contract for the agents that run the daily ingest, covering routing rules, page templates, and the exact maintenance sequence a run has to pass before it counts as done.

The public template, a generic version of this setup, is updated to match my setup and is live on GitHub if you want to try the structure yourself.

More soon as I keep evolving it.

Today marks one year since I started at Audible. This chapter is not really about that job. It is about a story I have known since I was a kid, a mindset I was already putting into practice before the layoff, and one that only accelerated once it hit.

The Fable

The ant works through the summer, carrying food back to the colony while the weather is good and nothing is threatening it. The grasshopper plays the guitar. When winter comes, the ant is fed and the grasshopper is not. Most people know the moral as "work hard, play later." I do not think that is quite it. The real point is that the ant does the unglamorous work when there is no visible threat and no one around it is doing the same. That is much harder than working hard during a crisis. Everyone works hard during a crisis.

I am one year into a job I like, at a company that treats me well. Layoffs are rarely something you see coming or control, which is exactly why this season, the one where nothing looks wrong, is the season the fable is about. This chapter is about what I am doing with it, and about the fact that I do not think I am doing enough of it yet.

Personal Equity

I use this phrase a lot now: personal equity. What I mean by it is a name and a credibility that belong to you, not to the organization you happen to work for.

Most professional visibility is hyphenated. People build a public voice, but it stays tied to their current title, their current company, the specific work they do there. That voice disappears the moment the job does. Personal equity is the opposite. It is what you would still have, standing in a room, if the company logo were not next to your name.

I started building mine two or three years before the Microsoft layoff, without knowing that was what I was doing. The real unlock came when I started writing publicly. Not polished essays at first, just honest notes about what I was learning, what was not working, how I was making sense of a piece of AI news. Someone once told me that you do not have to be somebody to say something. You have to say something to become somebody. I did not fully believe that until I lived it, and I am still building on it now.

The Grind

Here is what the ant work actually looks like for me right now, on a normal week.

I read and listen to AI news and research most days, not because anyone is asking me to, but because falling behind is the fastest way to lose an edge you cannot see slipping. I keep hobby projects running on the side, small things I build just to learn a tool or a technique before I need it for anything real. I take courses on topics adjacent to my actual job, not required by it. I speak at events and conferences whenever I can, and I advise a few startups, sometimes as a mentor, sometimes in a more formal board capacity. I host and show up for community events on weekends. Most weeks I am about as busy after work as I am during it.

None of this is because I think something bad is coming. It is because I do not want to find out the hard way again whether I was ready. The ant does not know exactly when winter will arrive either. It just keeps carrying.

Beyond the Job Title

One thing I have changed deliberately since landing at Audible is how narrowly I choose to define my own role.

The job title says product manager. But in the age of AI, I think staying inside the traditional boundaries of that title is its own kind of risk. I have pushed myself to go build evaluation systems, to sit with prototypes instead of only writing specs for them, to get into the data myself and look for the actual shape of where things are failing, instead of waiting for someone else’s summary of it. I do it because the version of a product manager who only manages the roadmap is a version that AI is going to make less necessary every year, and I would rather expand what I bring to the table than wait to see if that turns out to be true.

The Support System

People often think of a support system as something you receive. I think of it mostly as something you build by giving first, and giving more than you expect back.

That starts at home. It extends to the people around you at work and in your community. When you are in a season of comfort, the most useful thing you can do is make your time available to people who are not. I take a couple of calls most weeks with people navigating a layoff or a search, and I review resumes when asked. The other half of this is harder to say out loud. When you are the one struggling, you have to let people see it. You cannot always hold yourself together, and you should not have to. Keep the people close who will let you fall apart a little and stay anyway.

When the Arrows Do Not Fly Straight

Outside of work, I shoot competitively as an archer, and I coach other archers. Tournaments do not always go well. Some days your scores fall apart and nothing about your form feels right.

What I tell the archers I coach, and what I tell myself, is to trust the process when the arrows are not flying straight. Go back to what you practiced when nothing was on the line, and let that carry you through the moment when something is. The same thing was true during the search itself, on interview rounds that did not go the way I had prepared for, and it is still true now, in the ordinary setbacks that show up in any job, any project, any normal week. The discipline holds up, but only because I go back to the same fundamentals every time, not because any single round or setback goes perfectly.

Why I Build

People sometimes ask why I spend so much time running meetups, organizing hackathons, and investing in the Boston community. The honest answer is that I like being around people who build things, where creative energy is rampant and sharing and learning are appreciated traits.

It is also a deliberate, intentional, and consistent way to keep learning new skills and expanding my own horizons, alongside everyone else in the room doing the same. Layoffs are a normal part of this industry now. It is not a question of if, only when. Building a network and a habit of showing up for others, while you have the time and the standing to do it, is how you prepare for a moment you cannot predict.

What I Am Not Doing Well Enough

I want to end on something honest rather than something tidy, because I think that is more useful to whoever is reading this while things are hard.

I do not think I am evolving fast enough. I do not think I am learning as much as I should be, or tinkering and building as much as the moment actually calls for. The field is moving fast enough that standing still for even a season is a real cost, and most weeks I feel behind rather than ahead. The honest question I am sitting with right now is not whether the ant mindset works. It is whether I am doing enough of it, and how to get more efficient with the hours I have so that the equity I am building keeps pace with everything changing around it.

I do not have that fully solved. I am not sure anyone does. But I would rather tell you that plainly than pretend I have arrived somewhere I have not.

The Ant Mindset

A year ago I called myself de-hyphenated: someone standing in the gap between who I had been and who I was going to become next. I know now that the hyphen never fully goes away. You just learn to stand in it differently, less afraid, more useful to the next person standing there.

The garden I planted in the worst week of my professional life is still growing. I still go out and tend it, even on the days when nothing looks like it needs tending. The ant does not know when winter is coming. It just keeps grinding, all through the good weather, so that whenever winter comes, it is ready. That is the whole mindset. Not certainty. Just the carrying.

This is the fifth and final chapter of The Journey of Finding My Own Resilience. Thank you for reading this series, for your comments, and for the ways so many of you reached out along the way. If you are just joining now, the whole series is linked at the top of this piece.

I spent the past few days digging into where the term actually comes from, what Altman means by it now, and what the evidence for and against his claim looks like. The full piece is live on Forbes: Sam Altman Says We’re In The Singularity. What Does He Actually Mean?

A few things worth knowing before you click through.

The word carries two different meanings, and that’s the whole story. I.J. Good coined “intelligence explosion” in 1965: a machine that improves its own successor, repeatedly, without a human in the loop. Ray Kurzweil turned it into a public idea in his book “Singularity in Near” with actual dates attached, 2029 and 2045, dates you can check against reality. Altman’s own definition, from his 2025 essay “The Gentle Singularity,” is different. It describes how it feels when AI progress stops surprising you. One of these claims can be wrong. The other can’t.

Four people made a version of this claim in five months, and they don’t agree with each other. Jensen Huang says AGI is already behind us. Demis Hassabis says we’re in the foothills. Altman says we’re inside it. Musk says we’re past it and early. It’s four separate conversations sharing a word.

The evidence for acceleration is real. Frontier model releases at Anthropic and OpenAI have both compressed to about 60 days apart in 2026, down from roughly a year apart two years ago. Anthropic says Claude now writes more than 80% of the code merged into its own production codebase. None of that is disputed. What’s disputed is whether it adds up to a threshold crossed, or a curve that just keeps bending.

Altman may well be right about the direction. The piece walks through the math result that started this round of the debate, the security incident that got misread as evidence of machines setting their own goals, and OpenAI’s own roadmap for what would actually count as proof.

Read the full piece here: Forbes →

Twenty-five companies, including Microsoft, Nvidia, Meta, IBM, and Palantir, just told Washington not to regulate open-weight AI models out of existence. OpenAI, Anthropic, xAI, and Google didn’t sign. The split reveals two competing definitions of American AI leadership: frontier labs betting on the single smartest model, everyone else betting on the broadest, most competitive open ecosystem.

The fear isn’t hypothetical. Anthropic’s Claude Fable 5 was cut off globally for weeks by a Commerce Department directive, and when Hugging Face got breached by an OpenAI model in July, Fable refused to help clean it up, forcing Hugging Face to fall back on a Chinese open-weight model instead.

No rulebook substitutes for actually building the ecosystem.

Read the full piece on Forbes: https://www.forbes.com/sites/ashishbhatia/2026/07/25/america-cannot-regulate-its-way-into-building-a-dominant-ai-ecosystem/

That’s the wrong reaction.

What Anthropic just shipped is not a productivity feature. It is the opening move in a quiet but profound restructuring of how companies are built, staffed, and owned. We are watching the birth of a new economic paradigm.

The age of rented intelligence and digital employees.

This Is Not a Bot. It Is a Digital Employee.

Sure, it doesn’t feel like one yet. Here is what @Claude Tag actually does. It joins your Slack workspace. It reads your channels, in ambient mode, without being summoned. It learns your colleagues, your projects, your vocabulary, your workflows. It builds a graph of your organization from the inside. And when you need it, you tag it the same way you tag a teammate.

Fortune called it “a virtual employee that works within Slack.” Cat Wu, Anthropic’s head of product, described Claude Tag as “multiplayer”: one Claude instance per channel, visible to and usable by every member of your team. Andrej Karpathy, who just joined Anthropic, called it “a self-contained persistent asynchronous entity with org-wide tools and context working alongside teams of humans.”

A persistent entity. Working alongside your team.

This is not a chatbot you visit explicitly. This is a colleague who is always ambient and always working.

Already, inside Anthropic itself, Claude Tag is approving and incorporating 65% of the product team’s code changes. The legacy Slack integration gets retired August 3rd. Anthropic is at a $30 billion annualized revenue run rate, and as of May 2026, Anthropic has pulled ahead of OpenAI in enterprise adoption: 34.4% of U.S. businesses paying for Claude versus 32.3% for OpenAI, per Ramp’s AI Index.

The digital employee is already here, in your Slack channel, reading your messages, learning how you and your colleagues work.

From Explicit to Implicit Context

For the past few years, we have been in the era of explicit context. You give the model what it needs. You paste the document. You write the system prompt. You describe your company, your role, your situation. You upload the file. You build the RAG pipeline.

This is how most people still think about AI today: a powerful tool that needs to be fed.

Claude Tag changes that entirely.

When Claude is ambient inside your Slack, reading channels, following threads, observing who responds to what, noticing which decisions stick and which get relitigated, it is not waiting to be fed context. It is absorbing it. Passively. Continuously. From the natural flow of work.

This is implicit context. And it is a categorically different thing.

Explicit context is what you give a model in a session. Implicit context is what a model accumulates by being present where work actually gets done. One is episodic; the other compounds. One resets when the conversation ends; the other builds quietly for as long as the model is embedded.

The distinction sounds technical. The implications are not. Every explicit context system, every RAG pipeline, every system prompt, every carefully curated knowledge base, requires someone to decide what the model should know. Implicit context requires no such decision. The model simply observes. It is there when the VP of Sales messages the CEO at 11pm. It is there when engineering debates whether to ship. It is there for the judgment calls that never make it into a document but shape everything that follows.

That accumulation of implicit context over time is what I am calling Context Capital. And it is the reason this moment is different from every prior wave of enterprise software.

Human Capital to Context Capital

There is a concept that has organized corporate strategy for decades: human capital. The idea that people, their skills and knowledge and judgment, are a form of capital that companies accumulate and deploy.

We are entering the era of Context Capital.

Every AI model today is intelligent. That is table stakes, not the differentiator. What these models lack, what you spend hours constructing in system prompts and RAG pipelines and attached documents, is context. Specifically, your context. The context of your organization, your workflows, your decisions, your institutional memory, your way of thinking.

Context Capital is what that implicit context becomes when it accumulates inside your company over time. It is not just model access at a new surface. It is the model plus your data plus your workflows plus your conversations plus your decisions: the sum of everything Claude has observed without being explicitly told. It is the graph of your business that Claude Tag is quietly building.

Here is why this framing matters: human capital, at least in principle, can move. People leave. They take their knowledge with them. You can lose it and it takes time to rebuild.

Context Capital will not work that way.

Satya Nadella, CEO of Microsoft, published an essay recently that got 28 million views. His core argument: “AI models can continuously absorb the expertise of humans and organizations and commoditize it.” He introduced the concept of “token capital” alongside human capital. Nadella was warning enterprises to build their own. I am pointing at something different: who is building it, and who owns it when it is done.

Once Anthropic’s models have ingested two years of your Slack messages, your product specs, your customer conversations, your internal debates, your institutional memory, that context does not transfer. The model calibrated on your company, in your organization’s specific flavor, is not something you can export. You cannot rebuild it with a different provider overnight.

This is not model lock-in. It is context lock-in. And context lock-in is a far more durable form of dependency than any software subscription you have ever signed.

The Perpetual Learning Loop

Here is the mechanism that makes this different from every prior form of software lock-in.

Every time an employee tags @Claude in Slack, they generate what Nadella calls an “execution trace”: a record of the actual workflow, the actual decision, the actual outcome, inside a real company. In standard software, those traces are logged in cold storage for offline analysis and dashboards. In Anthropic’s model, they feed back into a learning loop.

The model is not just learning your data. It is learning your workflow, your judgment, your business niche.

It learns how your legal team frames risk. How your product team prioritizes features. How your sales team handles objections. The vocabulary your engineers use when something breaks. The way your CEO thinks about strategy. These are not data points. They are patterns of organizational cognition. They compound over time and form the bedrock of Context Capital.

Every quarter Claude Tag sits inside a company, the model grows more precisely calibrated to how that organization thinks and operates. After a year, it understands the company better than most new employees ever will. After two years, it holds institutional memory that no departing executive can take with them.

You are not building dependency on a tool. You are building dependency on a system that knows your organization the way no one else does. That is what makes the gravity of this almost impossible to escape once it sets in.

Analyst Kai Waehner calls this “data gravity”: the more context, fine-tuning, and institutional knowledge you invest in a platform, the harder exit becomes. He notes that agentic AI lock-in is “more durable than API lock-in because it accumulates at multiple layers simultaneously: the foundation model, the orchestration framework, the runtime environment, and the developer patterns that teams build around them.”

This is the perpetual learning loop Anthropic is building. No competitor can replicate it, because no competitor has your company’s two years of execution traces. The moat is not the model. The moat is the accumulated Context Capital.

The Trojan Horse

This move into Slack does not stand alone. Look at where else Claude is being embedded.

Claude is inside Excel. Inside PowerPoint. Inside Outlook. These are not peripheral surfaces. These are the places where the majority of knowledge work in the world actually happens.

Each of these integrations looks, on the surface, like a mutually beneficial partnership. Microsoft gets a better AI feature for its Office suite. Slack gets a more capable bot. Salesforce gets smarter automation. Anthropic gets distribution. Everyone wins.

That is the Trojan Horse.

The software partners think they are the host. They are not. They are the delivery mechanism.

When Claude is embedded in Excel, it does not just help users write formulas. It learns how enterprises model their businesses financially, what assumptions they make, what scenarios they stress-test, what decisions follow from the numbers. When it is in PowerPoint, it learns how companies tell their strategic story: what they emphasize, what they obscure, what narrative they use to align teams and court investors. When it is in Slack, it learns how companies actually communicate and operate.

Microsoft gets a better Copilot today. What Anthropic gets, indefinitely, is a window into how Microsoft’s enterprise customers think, decide, and operate.

The value exchange looks balanced at signing. It is deeply asymmetric over time. Anthropic’s side of the trade compounds. The partner’s side stays flat, or declines.

And here is the part that should concern every enterprise IT leader: most of this is not a deliberate decision. Employees are just using a feature. Nobody held a meeting to hand Anthropic their organizational intelligence. It accumulated silently, one @Claude tag at a time. As Waehner puts it, enterprises “are already making a default choice, and that default is usually determined by whichever vendor has the best marketing rather than the best governance posture.”

That is the sleepwalk. A thousand small decisions, none of them feeling significant.

As Anthropic’s layer matures inside these surfaces, the software beneath it starts to hollow out. Users stop needing to know Excel’s features when Claude models the data for them. They stop navigating PowerPoint’s layouts when Claude builds the deck. They stop searching Slack when Claude already knows what happened. The software partner does not disappear. It just stops being the product. It becomes the pipe.

And Anthropic, having spent years learning exactly how each surface is used, is better positioned than any of these partners to build what comes next, or to make what already exists redundant. This is what eating software from the inside actually means. Not hostile takeover. Not direct competition. A quiet migration of value up the stack, until the stack below becomes infrastructure and the intelligence layer above it becomes indispensable.

The Pricing Model Changes Everything

Think about what it costs to hire a human employee. There is a salary cap. You can negotiate it, you can raise it, but there is a ceiling. The employee has finite hours. They sleep. They leave.

Claude Tag has no ceiling.

The pricing model for AI agents is token-based and unbounded. There is always more work an agent can do. More threads to read, more documents to summarize, more code to write, more decisions to make. And every unit of work costs more tokens, which means more money to Anthropic.

The moment this model becomes indispensable, you have lost your pricing leverage. You cannot offboard the system that knows your business better than your own team. You cannot tell Anthropic the price is too high when its agents are woven into every workflow your people depend on.

Human capital had counterweights: unions, labor markets, portable skills, the ability to leave. Context Capital is arriving with none of those structures in place.

The company that owns the context owns the relationship. Anthropic is building that ownership one deployment, one workspace at a time.

Two Lock-ins, One Winner

If you are a product leader or an executive trying to make sense of this moment, here is the framework I keep coming back to.

There are two types of AI lock-in. Most people are worried about only one.

Model lock-in is when you build so much on one provider’s API that switching becomes expensive. This is real but manageable. Models are becoming more interchangeable. The cost of switching is going down.

Context lock-in is when the model you rely on has been attuned to your specific organizational context, your data, your history, your judgment. You cannot export context. It is, by definition, proprietary to the relationship between your company and Anthropic. And the perpetual learning loop means it deepens every single day you stay.

The test Nadella offers is stark: can you swap out the underlying model without losing the capability you have built? If the answer is no, you do not own your context. You rent it.

Claude Tag is a context lock-in play. The ambient reading, the org-wide knowledge graph, the persistent memory of your team’s work: all of it builds data gravity that Anthropic controls. And that leverage only grows.

The companies that navigate this well are the ones asking hard questions now. Who owns the data Claude Tag ingests? Can you export your company’s knowledge graph? What happens if Anthropic changes its pricing, its terms, or simply decides your industry vertical is one it wants to own directly?

We Are Sleepwalking Into This

I am not saying Anthropic is doing anything malicious. The incentives here are obvious and rational. Build the best AI. Embed it everywhere. Let the learning loop deepen. Charge for the dependence that follows. This is the calculus every enterprise needs to confront.

That is a sound strategy. It is also a strategy that, at scale, concentrates an extraordinary amount of economic power in a single company.

Satya Nadella reached for a historical parallel that I think is exactly right. When manufacturing was offshored in the late twentieth century, the GDP numbers looked fine. Corporate margins improved. Consumer prices fell. The structural damage, the gutting of industrial knowledge, supplier networks, engineering know-how built over generations, was invisible for years, until it wasn’t. Entire regional economies hollowed out. The expertise could not be rebuilt quickly once it was gone.

Nadella is arguing the same risk now applies to knowledge industries: law, finance, medicine, consulting, software. The quarterly earnings will look fine. Productivity will go up. Costs will go down. The structural damage, the quiet transfer of organizational intelligence to a small number of AI providers, will be invisible until it is complete.

Anthropic becomes the provider of Context Capital. We have not agreed on the vocabulary to describe it yet. We are still calling it a Slack feature.

The answer is not to refuse. The answer is to go in with open eyes.

Own your context. Demand portability. Ask your AI vendor whether your execution traces feed their learning loop or yours. Support open-source models and multi-provider strategies that give you real alternatives.

The digital employee has arrived. The question is who it actually works for.

Written in a personal capacity. I spent years at Microsoft building AI platforms, and I now work at Audible. This piece draws on public reporting only and these are my own observations unrelated to my employer.

Sources

1. Anthropic, “Introducing Claude Tag,” official product announcement. https://www.anthropic.com/news/introducing-claude-tag

2. Beatrice Nolan, “Anthropic launches Claude Tag, a tool that works like a virtual employee within Slack,” Fortune, June 23, 2026. https://fortune.com/2026/06/23/anthropic-claude-tag-virtual-employee-tool-slack/

3. Andrej Karpathy, public post on X, June 2026.

4. Anthropic $30 billion annualized revenue run rate, reported by VentureBeat, April 2026. https://venturebeat.com/technology/anthropic-says-it-hit-a-30-billion-revenue-run-rate-after-crazy-80x-growth

5. Ramp AI Index, May 2026. Enterprise adoption figures: Anthropic 34.4%, OpenAI 32.3%. https://ramp.com/leading-indicators/ai-index-may-2026

6. Satya Nadella, “Frontier Ecosystem,” June 14, 2026. https://snscratchpad.com/posts/frontier-ecosystem/

7. Kyle Belmonte, “Microsoft CEO Issues AI Warning,” TechTimes, June 15, 2026. https://www.techtimes.com/articles/318394/20260615/microsoft-ceo-issues-ai-warning-companies-that-rent-models-risk-industry-hollowing.htm

8. Kai Waehner, “Enterprise Agentic AI Landscape 2026: Trust, Flexibility, and Vendor Lock-in,” April 6, 2026. https://www.kai-waehner.de/blog/2026/04/06/enterprise-agentic-ai-landscape-2026-trust-flexibility-and-vendor-lock-in/

9. Dylan Anderson, “Anthropic Won the Market By Embedding AI, Not With a Better Model,” The Data Ecosystem (Substack), May 15, 2026. https://thedataecosystem.substack.com/p/anthropic-embedding-ai

10. Beatrice Nolan, “Renting AI from foreign providers is a national security risk,” Fortune, June 23, 2026. https://fortune.com/2026/06/23/cohere-ceo-sovereign-ai-anthropic-mythos-national-security-risk/

CIOs especially like that kind of predictability, a flat number on a budget line is how they like to see it. So far, AI pricing looked like it would behave the same way every other piece of enterprise software always had.

It will not. AI pricing is forking into two distinct models, and it is the enterprise usage, the half of the market that is growing the fastest right now, that is recalibrating to a new meter that has started to run. Consumers will keep their flat monthly tiers. Enterprises are heading somewhere else entirely: consumption-based pricing, billed by the token, governed like cloud infrastructure. The winners of the next phase will not be the companies that spend the most on frontier models. They will be the ones that route the right work to the right model and govern who gets to spend what.

On June 1, 2026, GitHub moved every Copilot plan to usage-based billing (source).

Here is the question worth sitting with: it is no longer “which model is best”. It is “what is this task actually worth?”

Two Markets, Two Pricing Models

Start with the part that makes this thesis credible: most consumers will stay away from the metered API billing, and for good reason. A flat monthly fee is predictable. Pay-as-you-go pricing can swing with usage, several multiples higher in a heavy week than in a light one, and that is exactly the kind of bill a consumer subscription is designed to avoid.

Both Anthropic and OpenAI sell flat monthly subscriptions to consumers. But they draw the line between “subscriber” and “heavy user” in very different places.

Anthropic guards the subscription. As of April 4, 2026, Claude Pro and Max subscribers can no longer route their plan limits through third-party agent frameworks like OpenClaw (source). The economics explain why. A single OpenClaw instance, run autonomously for a day, can consume $1,000 to $5,000 in API-equivalent compute against a $200-a-month Max plan (source). Anthropic estimated more than 135,000 such instances were running when it made the call (source). Heavy agentic use now gets routed to separate, pay-as-you-go billing. The flat tier survives only for Claude AI, Claude Code and Claude Cowork.

OpenAI has been more permissive so far, and its CFO will tell you why out loud. Sarah Friar put it this way: “If I was optimizing only for today, I would give every token to the API, order of magnitude more than to the consumer. However, we’re playing our own game. We have a strategy where we believe there’s an AI infrastructure layer, a utility like electricity, and in a future state, you’ll want to be able to serve the world at large: consumers, small businesses, large enterprises, governments” (source, ~31:08).

That is worth pausing on. The person who runs OpenAI’s P&L just told us, on the record, that the per-token money is in API consumption, and that the consumer subscription is a deliberate strategic subsidy, a choice, not the most profitable path available to her.

It also explains the funnel underneath it. OpenAI’s consumer “commitment curve” gives free users roughly seven prompts a day, Plus users about three times that, and Pro users roughly eleven times that (source, ~07:04). That is a tiered subscription structure engineered, on purpose, to grow a funnel.

Even a subscription has a meter. It just hides until an agent finds it.

The Meter Was Always Coming

Enterprise AI is becoming metered infrastructure, billed and governed the way you already bill and govern cloud.

GitHub is the cleanest evidence available right now. On June 1, it retired premium request units across every Copilot plan and moved everyone onto GitHub AI Credits, drawn down by token usage at roughly a cent per credit (source). Pro gets 1,500 credits a month. Pro+ gets 7,000. Enterprise gets 3,900 per seat. Need more than that? A new $100-a-month Copilot Max plan now exists for heavy agentic users.

Notice what stayed flat and what didn’t. Code completions (“Tab completions”), the small, constant suggestions as you type, remain included. The agentic, multi-step work, the part that plans, edits across files, and runs for minutes at a stretch, is what meters (source). That is not a coincidence. That is also the part developers love most, and the part that costs the most to run.

The backlash arrived within days. One developer on the $39-a-month Pro+ plan burned through 8% of a monthly allotment in two hours and projected the rest would be gone in under 48 (source). Thant’s a signal. Consumption pricing just made the real cost of AI visible to the people actually using it, for the first time.

Flat-rate enterprise AI was a subsidy. The subsidy is ending.

The “so what” is simple, and uncomfortable: AI spend stops being a tidy line item on a software budget and becomes a variable operating cost, the kind that moves with usage the way your cloud bill already does. It needs the same discipline you already bring to cloud and to data. Most enterprises are not bringing it yet. Two examples show exactly what happens when they don’t.

Uber Set the Budget on Fire

In December 2025, Uber rolled Claude Code out to its engineering organization and, to drive adoption, built an internal leaderboard ranking teams by AI usage.

It worked. By March, 84% of Uber’s roughly 5,000 engineers were classified as agentic coding users, and 70% of newly committed code traced back to AI assistance (source).

It also burned through Uber’s entire 2026 AI tools budget in four months, across Claude Code and Cursor, with monthly per-engineer costs running $500 to $2,000 (source). COO Andrew Macdonald called the disclosure a “head-exploding moment,” and was candid that the link between all that spend and customer value “is not there yet” (source). Uber has since capped per-employee spend at $1,500 a month per AI coding tool, with a request process for engineers who need to go higher (source).

They did not have a tooling or adoption problem. They had an incentive problem. A leaderboard rewards consumption. It does not reward outcomes. Uber built a machine that optimized for exactly the number that, six months later, set off alarms in finance.

Even Microsoft Pulled the Plug

Second proof point, and the one that should make every enterprise leader sit up.

Last December, Microsoft handed thousands of engineers Claude Code on the company’s dime. Six months later, it is cancelling those licenses across its Experiences and Devices division, the group that builds Windows, Microsoft 365, Teams, and Surface, and pushing those engineers onto GitHub Copilot CLI by June 30 (source).

Two things are going on here, and they reinforce each other. It is cost: the move lands right at the close of Microsoft’s fiscal year, and Claude Code’s enterprise pricing scales with usage in exactly the way that just blew up Uber’s budget (source). And it is platform control: Microsoft owns GitHub, sells Copilot to the rest of the world, and cannot keep doing that credibly while thousands of its own engineers quietly prefer somebody else’s tool, calling Claude Code “perhaps a little too popular” internally (source).

Sit with the paradox here. Engineers adopted Claude Code so fast, and so completely, that it took an active, top-down policy decision to pull it back. Demand was never the problem. Ungoverned demand was.

When the company that owns the platform decides it cannot afford its own meter, that is a signal. And it is aimed directly at every enterprise still treating AI spend as a fixed cost expense.

The Pattern Is the Point

Uber. Microsoft. GitHub’s own meter, switched on for everyone at once. Three different stories. One pattern: unmetered access to a frontier model does not scale, no matter how sophisticated the company running it is or how deep its pockets are.

The root cause is a habit, and it is a habit almost every team has. We send every prompt to the best model available. A JSON formatting check, a one-line label, a two-sentence summary, all of it routed to a top-tier reasoning model that can cost an order of magnitude more per token than a smaller model built for exactly that job. Industry pricing data from this year shows the gap is not small: frontier and budget-tier models can differ by 10 to 30 times in cost per token for comparable work, and the full market spans a roughly 600-times range from cheapest to most expensive (source).

The frontier itself is getting dramatically cheaper, on its own, without you doing anything. Friar puts the cost drop from GPT-4 to GPT-4o at roughly 97% over about two years, and notes that even the newer GPT-5.5, despite a 2x increase in list price, still lands customers a 20% to 30% lower cost per token, because it is so much more efficient at the same task (source, ~17:51).

Sit with that for a second. If a single model line gets 97% cheaper in two years, then defaulting every task, forever, to the most expensive model on the menu is not a productivity strategy. Most enterprise AI traffic is routine. We are paying frontier prices for clerical work, on model and hardware that keeps getting cheaper for exactly the work we are not bothering to route.

So what do you actually do? You stop treating every request as equal.

Route the Work, Govern the Spend

This is the change in the operating model and the mindset, and it is also the only sustainable operating model on this side of the fork. Think of it as a loop, not a one-time fix:

1. Classify the task. Routine work (formatting, extraction, classification, short summaries) is not the same as genuine reasoning (architecture, multi-step planning, novel synthesis). Most of what flows through an enterprise is the first kind, especially in the knowledge work domain.

2. Route to the cheapest sufficient model. Send the routine work to a small or mid-size model. Reserve the frontier for the tasks that actually earn it. Cascade between tiers, fall back when a smaller model stalls, and cache what repeats. This is not a theory. Researchers at UC Berkeley built exactly this kind of router, RouteLLM, open-sourced it, and found it could send most everyday queries to cheaper models while preserving 95% of GPT-4’s quality, at a cost reduction of more than 85% on standard benchmarks (source). Vendors selling the same idea into enterprises report similar results in production: teams that route consistently cut AI spend by 60% to 80% with no real hit to quality, because most of what they were sending to the expensive model never needed it (source).

3. Set role-based token quotas. A staff engineer working a critical path is not the same as everyone running agents on everything, all day. Allocate budget by role, by team, by use case, the way you already allocate cloud spend.

4. Make spend visible. Per-team dashboards. Attribution down to the workload. Anomaly alerts before the bill arrives, not after. This is FinOps for AI, and it is already showing up in procurement decisions, not just blog posts. In June 2026, the cybersecurity firm BeyondTrust adopted a governance platform purpose-built to tag and track Anthropic token spend the same way it already tracks cloud spend, citing the need for organizational structure and compliance reporting as AI usage scaled across the company (source). At the volumes we are now talking about, this stops being optional.

5. Measure value, not usage. This is Uber’s mistake, compressed into one line. Reward outcomes. Never reward a leaderboard.

The answer to runaway cost was never to ban the tools. Engineers clearly want them, and for good reason: the tools work. The answer is routing and governance, applied with the same seriousness you already bring to every other meter your company watches.

Build the Routing Layer Before the Bill Teaches You To

The first wave of enterprise AI was about access. Get the best model into everyone’s hands, as fast as possible, and trust that output would follow. It did, for a while. Then the bill arrived.

The next wave is about discipline. Route the right work to the right model. Govern who spends what, and why. Measure the outcome, not the volume.

The meter is already on. The only open question left is whether you build the routing layer now, on your own terms, or whether you build it later, in a hurry, the way Uber and Microsoft just did.

Written in a personal capacity. I spent years at Microsoft building AI tools for Copilot Studio, and I now work at Audible. This piece draws on public reporting only and these are my own observations unrelated to my employer.

Sources

1. GitHub, “Updates to GitHub Copilot billing and plans,” GitHub Changelog, June 1, 2026: https://github.blog/changelog/2026-06-01-updates-to-github-copilot-billing-and-plans/

2. GitHub, “GitHub Copilot in Visual Studio Code: May releases,” GitHub Changelog, June 3, 2026: https://github.blog/changelog/2026-06-03-github-copilot-in-visual-studio-code-may-releases/

3. “’What a joke’: GitHub Copilot’s new token-based billing spurs consternation among devs,” TechCrunch, May 30, 2026: https://techcrunch.com/2026/05/30/what-a-joke-github-copilots-new-token-based-billing-spurs-consternation-among-devs/

4. “Anthropic blocks OpenClaw from Claude subscriptions in cost crackdown,” The Next Web: https://thenextweb.com/news/anthropic-openclaw-claude-subscription-ban-cost

5. “Anthropic cuts off the ability to use Claude subscriptions with OpenClaw and third-party AI agents,” VentureBeat: https://venturebeat.com/technology/anthropic-cuts-off-the-ability-to-use-claude-subscriptions-with-openclaw-and

6. “Anthropic temporarily banned OpenClaw’s creator from accessing Claude,” TechCrunch, April 10, 2026: https://techcrunch.com/2026/04/10/anthropic-temporarily-banned-openclaws-creator-from-accessing-claude/

7. “Uber burned through its entire 2026 AI budget in four months,” Yahoo Finance: https://finance.yahoo.com/sectors/technology/articles/uber-burned-entire-2026-ai-180347400.html

8. “Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it’s worth it,” Fortune, May 26, 2026: https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/

9. “Uber Caps Usage of AI Tools Like Claude Code to Manage Costs,” Bloomberg, June 2, 2026: https://www.bloomberg.com/news/articles/2026-06-02/uber-caps-usage-of-ai-tools-like-claude-code-to-cut-costs

10. “Uber caps employee AI spending after blowing through budget in four months,” TechCrunch, June 2, 2026: https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/

11. The Verge (Tom Warren, “Notepad”), on Microsoft cancelling internal Claude Code licenses across its Experiences + Devices division: https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad

12. Sarah Friar, “OpenAI CFO Sarah Friar: IPO, AI Rivalries, New Device, and Spending $100B+ on Compute,” YouTube interview (commitment-curve figures ~07:04, cost-deflation figures ~17:51, “every token to the API” ~31:08):

13. “LLM API Pricing Comparison In 2026: Every Major Model, Ranked By Cost,” CloudZero: https://www.cloudzero.com/blog/llm-api-pricing-comparison/

14. “Why Most Teams Overpay 40-85% for AI: The Routing Cost Math,” MindStudio: https://www.mindstudio.ai/blog/best-ai-model-routers-multi-provider-llm-cost-011e6

15. LMSYS Org, “RouteLLM: An Open-Source Framework for Cost-Effective LLM Routing” (UC Berkeley / Anyscale / Canva research, presented at ICLR 2025): https://www.lmsys.org/blog/2024-07-01-routellm/

16. “Kion Advances FinOps+ With Anthropic Token Spend Management and Automated Governance Controls” (BeyondTrust adoption), Business Wire, June 4, 2026: https://www.businesswire.com/news/home/20260604919308/en/Kion-Advances-FinOps-With-Anthropic-Token-Spend-Management-and-Automated-Governance-Controls

This is not a story about AI hype fading. I am an AI believer, AI is working and the adoption is real. The problem is that the way enterprises adopted it, by treating unlimited consumption as a virtue, has created a spending crisis that most organizations were not prepared for. The first phase was enablement. The next phase is governance.

The transition has a name. Let's call it Tokenomics.

First, Let's Talk About Tokenmaxing

Before you can govern AI spend, you have to understand how it got out of control. The answer is a behavior called tokenmaxxing (source). The practice has "turned token usage into a benchmark for productivity and a competitive measure of who is most AI-native."

The trigger was internal leaderboards. Inside Meta, an employee built a ranking system called "Claudeonomics" that scored all 85,000+ employees by token consumption and gave titles like "Session Immortal" and "Token Legend" to the top users (source). The result? Meta's engineers burned through 60.2 trillion tokens in a single month. At Anthropic's API prices, that would cost $900 million. Even at enterprise discounts, the bill likely landed above $100 million. For a month of token consumption.

Microsoft ran the same playbook. An internal token dashboard, live since January 2026, ranked employees by usage across the whole company. Distinguished-level engineers, the most senior individual contributors, landed in the top five. VP-level leaders made the top ten and twenty, despite spending most of their days in meetings. The number had become the signal of being "AI-native," so people optimized for the number.

The one counterexample worth naming is Shopify. The company built the first internal token usage dashboard, then renamed it from "leaderboard" to "usage dashboard" when they saw where the competitive framing was going. They added circuit breakers that automatically cut off access when personal spend spiked in a single day, which caught runaway agents and surfaced infrastructure bugs. And the head of engineering personally reviews what top-spending engineers are actually building (source). The result is high usage that connects to real work. That is the standard every org should be building toward.

Why did tokenmaxxing take hold so widely? Because in the enablement phase, consumption was the metric. Adoption dashboards tracked active users, message volume, and seats deployed. More tokens meant more engagement, and more engagement meant the rollout was working. The behavior the org rewarded was the behavior it got. And now it is the source of the problem.

The Budget Wall

The bill is no longer theoretical. Enterprises are hitting it in real time, and some of the most sophisticated technology companies in the world are among the first to feel it.

Uber rolled Claude Code out to its roughly 5,000 engineers in December 2025 and, to drive adoption, built an internal leaderboard ranking teams by AI usage. It worked spectacularly. By March, 84% of engineers were classified as agentic coding users, and 70% of new code traced back to AI assistance (source).

It also burned through Uber's entire 2026 AI tools budget in four months, at per-engineer monthly costs of $500 to $2,000 (source). COO Andrew Macdonald called it a "head-exploding moment" and admitted, publicly, that the connection between all that spend and customer value "is not there yet" (source). Uber has since capped per-employee AI spend at $1,500 a month and built a request process for engineers who need to go higher (source).

Microsoft followed a similar arc. In December, it gave thousands of engineers Claude Code on the company's dime. Six months later, it cancelled those licenses across its Experiences and Devices division, the group that builds Windows, Microsoft 365, Teams, and Surface, and migrated engineers to GitHub Copilot CLI by June 30 (source). Part of that decision is about platform control, Microsoft owns GitHub and sells Copilot to the world. But cost is also named plainly. Claude Code's consumption-based pricing scales exactly the way that just broke Uber's budget.

The pattern is not limited to tech giants. Just recently, I was talking to a friend who works at a leading financial institution. He told me they had been struggling with their Claude API spend to the point where they had to make a hard call: pull all employee access to Claude Opus and downgrade everyone to Sonnet and Haiku. Not because the Opus outputs weren't good. Because the budget was gone. An entire tier of model capability, taken off the table, not for performance reasons but for cost ones.

The ROI Question Is Not Going Away

Here is the data that should make every executive uncomfortable. MIT research found that 95% of generative AI pilots delivered no measurable profit-and-loss impact despite billions invested (source). Enterprise AI token consumption has increased 13x since January 2025 (source). And 73% of enterprises reported their AI costs exceeded original projections, even as the blended cost per million tokens dropped 67% year-over-year (source).

Read that last one carefully. Prices are falling. Spend is rising faster. Total cost equals price per unit multiplied by volume consumed. The first variable is moving in your favor. The second is moving against you, faster.

Average enterprise AI spend is projected to jump roughly 65%, from $7 million in 2025 to $11.6 million in 2026 (source). For companies that built no governance layer during the enablement phase, that number can arrive as a surprise.

The conversation in every CFO and CIO meeting right now is the same: we are spending more than we budgeted. We cannot point to the return. And we cannot get in front of the finance team with nothing but a usage dashboard.

Enterprises Need Tokenomics

Tokenomics is not cost-cutting. It is not a mandate to do less with AI. It is a discipline: the principles and practices for matching AI spend to AI value, at the organizational level.

Think of what cloud FinOps did for infrastructure spend. Before FinOps, cloud bills were treated like electricity, a utility cost you paid and moved on. FinOps brought attribution, allocation, tagging, anomaly detection, and optimization into the operating model. It did not reduce usage. It made usage visible, accountable, and sustainable. That is what Tokenomics does for AI.

Here are some ideas:

• Route to the right model. Not every task requires your most capable model. A JSON formatting check, a label classification, a two-line summary: these are not frontier tasks. They are clerical tasks, and smaller models handle them well at a fraction of the cost. The gap between frontier and budget-tier models can be 10 to 30 times in cost per token for comparable routine work (source). Researchers at UC Berkeley built RouteLLM, an open-source routing framework that sends everyday queries to cheaper models while preserving 95% of GPT-4-level quality. Cost reduction on standard benchmarks: over 85% (source). Teams that route consistently report 60% to 80% cost reduction with no meaningful quality loss (source).

• Implement role-based model access. Not everyone in an organization needs the same model tier. A data scientist building complex multi-step reasoning pipelines has different needs than a customer service rep drafting email responses. Define model tiers by role. Give your power users access to the frontier. Give everyone else the model that matches the complexity of their actual work. This is not about restricting access for its own sake. It is about allocating the right capability to the right job, the same way you think about software licenses or cloud instance types.

• Set token quotas for roles and business functions. Token budgets work exactly the way cloud resource budgets work. Define an expected monthly allocation per role, per team, per use case. Build alerts that fire at 70% consumption so you are never surprised at month-end. Build a request process for exceptions, the way Uber eventually did, but proactively rather than reactively. The organizations that will manage this best are the ones that treat token budgets as a first-class planning artifact, not a dashboard you check after the fact.

• Renegotiate your AI contracts. Most enterprise AI contracts were signed during the enablement phase, when the primary goal was access, and the pricing was frontier-model-first, volume-light. That was the right deal for year one. It is not the right deal for year three. Go back to your AI vendors with actual consumption data, tiered usage patterns, and a conversation about committed spend versus on-demand. AI labs are competing aggressively for enterprise contracts right now, and you have more leverage than you did eighteen months ago.

• Explore cloud-hosted open source models. The open-source model ecosystem has matured considerably. Meta's Llama family, Mistral, Qwen, and a growing roster of specialized models are now available through AWS Bedrock, Azure AI, and Google Vertex, alongside proprietary options. For workloads where the task is well-defined, a cloud-hosted open-source model can deliver strong results at meaningfully lower cost than frontier proprietary APIs, without the operational overhead of running your own infrastructure.

• Consider self-hosted open source for high-volume, constrained workloads. For organizations with the engineering capacity to support it, self-hosted open-source models change the economics entirely. No per-token billing. Predictable infrastructure costs. Full control over the data. This is not the right answer for all workloads, and it is not free to operate. But for the right use cases, specifically high-volume, well-scoped tasks where a smaller model clearly fits, the cost differential versus a frontier API is substantial. It is worth putting into the conversation.

OpenAI May Be Doing the Math as well

There is news worth watching here. The Wall Street Journal reported in June 2026 that OpenAI is mulling sharp cuts to what it charges for tokens, in anticipation of similar moves from Anthropic, as both companies compete aggressively for enterprise market share (source).

This is not entirely surprising. API prices across the major providers have already dropped roughly 80% since early 2025, driven by hardware improvements, model efficiency gains, and intensifying competition (source). OpenAI CFO Sarah Friar put the cost drop from GPT-4 to GPT-4o alone at 97% over roughly two years (source). The direction of travel on unit economics is clear.

The strategic logic behind this price reduction is clear: it is undercutting Anthropic's lead with Claude Code and Opus-class models. OpenAI is not optimizing for this month's revenue. They are optimizing for the enterprise footprint and long term value, when AI spend is embedded into operating models and switching costs are real. Lower token costs today buy stickiness. That is a lifetime value calculation, not a margin one.

If the rumors are right and prices do fall significantly, it will provide some relief to budgets under pressure. But it will not solve the governance problem. If you have no routing layer, no quota system, and no attribution model today, lower token prices will simply mean the same ungoverned consumption at a lower unit rate. Governance is not a response to high prices. It is how you run a mature AI program.

Tokenomics Is Not What the AI Labs Want

Here is the uncomfortable truth: Tokenomics is not in the interest of the companies selling tokens. AI labs are consumption businesses. Their revenue model runs on volume. Every routing decision that sends a task to a cheaper model instead of a frontier one, every quota that limits agent loops, every self-hosted deployment that replaces an API call, is margin they do not collect.

The incentive to build you a great cost control layer is not aligned with the incentive to grow your spend. This is where the opportunity lies. For startups, building the governance and FinOps layer for enterprise AI is a real market, and it is early. Companies like Kion are already moving into AI token spend management the way cloud cost management tools did a decade ago (source). There is a generation of tooling to be built: token routers, quota managers, attribution dashboards, cost anomaly detectors, ROI calculators that connect spend to business outcomes.

For enterprises, the leverage is real too. You have consumption data. You have contract renewal cycles. You have the ability to move workloads. Demand better governance primitives from your AI vendors natively, not just through third-party wrappers. Ask for cost attribution by team and use case, anomaly alerting, and tiered access controls as first-class platform features, not add-ons.

The labs will follow the enterprise money. They will have to do it.

The enablement phase built the capability. The Tokenomics phase is where you build the discipline to sustain it.

Written in a personal capacity. I spent years at Microsoft building AI tools for Copilot Studio, and I now work at Audible. This piece draws on public reporting only and these are my own observations unrelated to my employer.

The Pipeline Builds

The Audible conversations had started about five weeks in.

The role was deeply entrenched in generative AI platform and AI evals. I had spent years building adjacent to exactly this problem. I my profile was very aligned the moment I read the job description.

There was one honest thing I had to deal with first: I was not an Audible listener. I had tried the platform years earlier. If I was going to interview for a role at the world's leading audiobook platform, I needed to understand it from the inside, not just describe it from the outside.

The interview came with a month of free Audible subscription. I signed up. I started listening on my commutes and evening walks. I spent time understanding the platform, the catalog, the curation, the way it fits into the rhythm of a day. By the time I walked into the first round, the familiarity was real. I had earned it.

Other conversations were running in parallel. The pipeline had taken several weeks to build, and it was finally moving across multiple tracks at once.

The Call in the Car

The first round result came in while I was in the car with my cousin and her family. My phone rang. The recruiter told me the hiring manager had really liked me and wanted to move forward.

Everyone in the car started celebrating before I had even finished the call.

I kept saying it was only the first round. That nothing was decided yet. They kept celebrating anyway. And somewhere in that noise, I felt relived too.

The Last Round

The eighth round was with the head of AI.

I had spent eleven years in the AI space and I knew the material. But there is something different about sitting across from the person who leads the function you are hoping to join, going deep on the exact topics you care most about. That conversation carries a different weight than any other in the loop.

We went deep on AI evaluation. I answered every question. What I was not sure about walking out was the conviction with which I had answered them. I had set a high benchmark for myself going in, and I was not certain I had met it. When you are that close to the end of something long, every small thing feels like it could be the one that tips things over. I was stressed throughout the day, I remember calling Sachin, and saying my might have messed it all up on this last round.

My wife told me one difficult round would not undo other good ones. She was right and I knew she was right. But I did not sleep well that night.

Months later, after I had joined, I told him that conversation had scared the hell out of me. He laughed and said I had been fine. That he had been convinced. I wished I had known that walking out of that interview, so I wouldn’t have freaked out the whole day after the last conversation.

The Offer

At the very start of the search, a friend had told me I would have more than one offer to choose from when the time came. I remember thinking that was an optimistic thing to say.

Over the course of the search, four offers came in.

The first arrived on day fifty-nine, just a day before the Microsoft grace period expires.

I was in my home office downstairs, preparing for yet another interview, when the call came. The recruiter told me Audible was going to make an offer. I put the phone down and sat there for a moment. The weight that had been sitting on my chest for months lifted all at once. I fist-bumped the air. Then I almost slumped in the chair with relief.

I went upstairs and told my wife and son that Audible was going to make an offer.

The role had initially been based out of Newark. I would have had to move the family. But Audible offered me the Cambridge office instead. That detail, the fact that they made room for where my life was, said something about the kind of company it was before I had even started.

My wife and son hugged me. We went out to eat that evening to celebrate.

Having four offers to compare was something I had not fully prepared myself to feel. Each one was a validation on its own. My friend had been right: there was more than one. Together they told me something I had needed to hear for months, that the work I had put in had actually worked, that the nineteen-year gap had been closed, that I still had it.

All four roles were AI-centric, and there was real parity among them. But the choice was not difficult. I love books. I had spent three weeks becoming an Audible listener specifically to earn my way into those interviews. The alignment was exact: generative AI, conversational discovery, products at scale for people who love stories. It was the work I had spent the better part of a decade building toward.

The Transfer

Landing the offer was not the end of the immigration process. It was the beginning of a new, slower one.

After accepting, the visa transfer papers go to the new company's immigration lawyers. The process takes three to four weeks. During that time, the 120-day clock is still ticking. You have an offer. You know where you are going. You know who you will be working for. But you cannot start until the transfer clears, and the system takes its own time regardless of how ready you are.

There is not much you can do during those weeks except stay close to the process, respond to every request immediately, and wait. I did all of that. And then I waited.

It cleared. I started at Audible in August.

None of that happened because things fell into place on their own. It happened because I kept moving. Because my family held me together, when I was falling apart. Because my prep partner kept the sessions going. Because strangers on LinkedIn offered their networks and support. Because my friends kept checking in. Your support system is as critical in this process as your own competency.

The Harvest

That summer, the saplings I had planted in May grew into something.

I had sat on the deck in those early days wondering if I would still be there to see them. I was.

The garden thrived and delivered exactly what it had promised. The plants I had put in the ground during the hardest week of my professional life grew through the spring and the summer and came up full. I was out in the backyard in August picking what I had planted, and I stood there for a moment just to feel the gratitude for everything I had in that moment.

Plant in the hardest season. Show up every day. Trust the process even when you cannot see the outcome. Be there for the harvest.

Next is Chapter 5. The search is over but the story is not. It is about the hyphen, the space between who you were and who you are becoming

Nineteen Years

I had not done an external interview in nineteen years.

I joined Nokia as a direct hire, and spent eight years there before Microsoft acquired the company. I moved with the organization, and spent eleven years at Microsoft after that. Every role I held had come through internal transitions. I had never had to learn how companies actually hire at senior levels today.

That said, I had not been idle. In the weeks after the layoff, I had put real time into product sense frameworks, behavioral prep content, interview walkthroughs. I thought I had a handle on what was coming.

I started applying early, some of the most competitive AI product roles in the market. The screening calls were fine. Talking about my work has never been the hard part. But when the full rounds started, I realized quickly that watching someone else do something and actually doing it yourself under pressure are two completely different things. The early rejections taught me that. They were the warm-up games before I understood what the big league actually required.

The Rules Had Changed

Around week three, an Amazon recruiter called. Her name was Pooja Motiramani, and she told me what I needed to hear.

My resume was too generic. My approach was too broad. I was casting the net at everything and landing nothing. She was direct about it: "If you do not follow the structure, you will not succeed. But if you follow it, you have a good chance. Your profile is well aligned."

I was a scatter plot before that call. I realized it the moment she said it.

The modern senior PM interview has a shape. Amazon's is built around Leadership Principles and the STAR framework: Situation, Task, Action, Result. Every answer has a structure. Every story needs to thread through specific values. If you do not know the shape, you are speaking a language the panel cannot hear, no matter how strong your experience actually is.

I stopped applying cold the same day. Warm referrals only, introductions from people who could put a name to my application. A cold application at senior levels is mostly a wasted effort. With a referral, you are a person. Without one, you are a document in a pile.

The Sunday Note

A few days after the rejections, I got a message on LinkedIn from someone named Sejal Sagar. I did not know her. She was connected to Dilyana Ocetova, a friend of mine in Boston, had seen the announcement, and reached out to ask if she could help.

We got on a call the next morning. She had spent the weekend preparing for it.

She walked me through a list of second-degree contacts at companies hiring AI PMs at my level, with notes on each. Google Alert queries to surface new openings as they came out. How to make a resume ATS-compliant so it survives automated screening before a human ever sees it. And one rule she repeated: never apply without a referral. If you cannot find one, reach out to me and I will find the connection for you.

This was a Sunday. She had put all of that together on a Saturday for someone she had never met.

That is what I mean when I say the right people appeared at the right time.

Around the same time, a Ana Auz-Law friend and colleague from Microsoft connected me with Bill Heilmann, who went through my LinkedIn profile from top to bottom and rebuilt it: structure, keywords, visibility. All of it optimized for how recruiters actually search. The scatter plot Pooja had described was slowly starting to look like a target.

The Alliance

The same week, I got in touch with Sachin Mathur. We had been colleagues at Microsoft, and when the announcement came we connected almost immediately. There was no formal conversation about it. We just understood that we were going to do this together and hold each other to it.

What that turned into over the months that followed was a real daily partnership. Standups to check in on where we each were in our pipelines. Mock interviews where we played both sides, then went through every detail in the debrief afterward. Texts after rounds, working out together what had landed and what had not. Strategy conversations on how to position experience for roles we had not held before.

When I hit system design rounds, Sachin ran those sessions. He understood exactly what was being tested and prepared me the right way. But that was only one piece of it. He was the person I called after the hard rounds. The person I worked through difficult follow-up emails with. We were each other's first audience, toughest critic, and most reliable source of honesty.

He was running his own search at the same time. We both knew that the way to stay in it was to stay locked in together.

The Work

With a system in place, I went deep on the prep itself.

I built out twenty-five behavioral interview situations. Not talking points. Full stories, pulled from actual moments in my career: the difficult calls, the failures, the times I had been wrong and had to change course, the times I had held a position under pressure. Some of them I had not thought about in years. Digging them out was uncomfortable in a useful way. It forced me to be honest about who I actually am when things are hard, not who I wished I had been.

I practiced them on morning walks with my son. He would give me a scenario and I would walk him through the answer.

His feedback was simple. Dad, your stories are good. But tell them like you mean it. Watch the pauses and the filler words. And when you set up the situation, make sure I can follow it before you get into what you did.

He did not know anything about PM interviews. But he had heard me talk about this work his whole life. The same person who had told me on the worst evening that he had zero doubt I would land was now telling me on morning walks to slow down and believe it. I took that seriously.

For delivery, I used ChatGPT voice mode as a sparring partner. I would run through a scenario and listen back. The gap between what I thought I was saying and what was actually coming across was real, and I needed to hear it out loud.

Somewhere in the middle of all of it, the prep started to feel different. I started looking forward to it. Not because the stakes were lower. Because I was getting better, and I could feel it.

Five Hundred

By the time the first offer came, I had submitted close to five hundred applications. That is the actual number.

Less than two percent came back with any response. Referrals converted at around thirty percent, which is why every warm introduction mattered and why I followed Sejal's advice and stopped applying cold. Recruiter and hiring manager calls had roughly a fifteen percent chance of turning into a full round. Only a fraction of those became offers.

I tracked all of it. I treated the search the way I would have treated a product launch: what was moving, what was stalled, where to put energy. Keep creating opportunities. That was the rule I repeated every morning.

I was interviewing across domains I had never worked in before: fintech, edtech, healthcare, cybersecurity, gaming. Each one required me to go deep on the company, understand what they were building, learn what mattered to that specific team. Every conversation was its own world.

The number that surprised me most was not the rejections. It was how broad the domains were. I had spent over a decade in enterprise software and AI. But the search showed me that the skills I had built were more transferable than I had realized. I was being taken seriously in places I would never have considered eighteen months earlier.

That was the rebound. Not the offer. The rebound was the moment I understood I could compete in all of those rooms.

The pipeline was still running. I was still in it every day. What I did not know yet was that fifty-nine days in, it would all start to land.

In the next chapter, the pipeline starts to deliver. A phone call comes in while I am in the car with my family. Offers follow. And the clock starts again, for a completely different reason.

120 Days

I am on an H-1B visa. For anyone unfamiliar with what that means in a layoff: the moment your employer-sponsored job disappears, a clock starts ticking. In my case, two windows combined to give me my total runway:

60 days of continuation pay from Microsoft, during which the H-1B remained valid.

60 days of standard H-1B grace period after employment ends.

120 days total. That was it. That was everything.

And the clock was not the only thing narrowing the field.

Around two-thirds of the reach outs I made in those early weeks came back as immediate rejections, not because of my experience or my background, but because of my visa status. H-1B sponsorship is a filter many companies will not pass through. I would reach out, a conversation would begin, and then: "Unfortunately we are not able to support visa sponsorship at this time." Over and over.

The clock was running. And two of every three doors were locked before I could knock.

The day after the announcement, I put a post-it on the wall above my home desk. I wrote the number 120 on it. Every single morning after that, I crossed out the previous number and wrote the new one. That post-it became the most honest object in my life. You can still numbers scratched out with force, new ones written underneath with something between determination and fear.

My goal was not to use all 120 days. My goal was to land a job before day 60. Because here is something people outside this system do not always understand: a visa transfer does not begin the moment you start applying. It begins the moment a new employer makes you an offer and agrees to sponsor your visa. Until that moment, everything is in legal suspension. You are not transferred. You are not protected. You are waiting.

I had watched this play out for someone I knew. He had received an offer. He was waiting for the transfer to process. Then a government shutdown hit. The paperwork stalled. His company rescinded the offer. He had to leave the country with his family.

That story lived inside me every day of those 120.

The Saplings

It was May. New England spring was doing what it always does: erupting. Light green everywhere. Trees bursting. Flowers pushing through. The whole world outside my window was insisting on new life.

Earlier that week, I had bought saplings to plant in the garden. We always start planting in May in New England. Anything earlier risks a late frost that kills everything overnight. Those plants were meant to be harvested two to three months later, right around August.

One afternoon that first week, I sat on my deck at lunchtime. I was not hungry. I had a glass of water and some snack. I looked at those saplings, small and fresh and full of potential, and I thought: I might not be here to see them grow.

If I could not find a job in time, we would be selling everything. The furniture. The belongings. The life we had carefully assembled around us in this city. We would be leaving for a country my son had not called home in years, pulling him out of the school and friendships that made up his entire world. My wife would be leaving the career she had built here.

All of it, gone. Because of a sixty-day clock.

The garden was beautiful that spring. I sat with that beauty and let it break my heart a little, because grief and hope can live in the same moment if you let them.

The Routine That Saved Me

I made a decision in that first week: I would treat the job search like a job. Same hours. Same discipline. Same accountability.

I was at my home office by 9 every morning. Two hours on job boards, applications, recruiter outreach. Midday: research on target companies, online courses, interview prep. Afternoons: refining my resume based on feedback, practicing answers, reading about the companies I was targeting. I used AI as a constant companion on walks and commutes, talking through scenarios out loud, asking myself hard questions, rehearsing how to reason clearly under pressure.

In the evenings, my wife and son and I took walks together. Every evening, without fail. It was the thing that kept me from disappearing into the anxiety. Family was the anchor when everything else was drifting.

I also built accountability into the search itself. A small cohort of friends and former colleagues in similar situations, people who understood exactly what was at stake. We ran daily standups. We did mock interviews and debriefs. We exchanged frameworks for how to answer different types of questions. We shared what was working and what was not. I do not think I would have made it without them.

And every morning, I changed the number on the post-it.

Part of that structure became something I had not planned. I started running Friday office hours. An open hour, every week, for anyone else who had been laid off and was navigating the same territory. What I was figuring out in real time, how to approach an interview after nineteen years, how to read a company’s signals, how to pace yourself without losing yourself. I shared with whoever showed up. Some weeks three people, some weeks one, some weeks more.

But showing up for them made me show up for myself. Giving back what I was getting from others kept me anchored. It reminded me that I still had something to offer, even in the weeks when I was not sure I did.

The Day of Three Rejections

The rejections came throughout the search, as they always do. You learn to process them. You get better at separating the pain of rejection from the meaning of it.

But then came one particular afternoon.

Three rejections arrived in sequence: Microsoft AI, Wayfair, and Apple. By late afternoon, all three had landed in my inbox.

The Microsoft AI and Wayfair interviews had gone well. I had left both conversations feeling genuinely hopeful. I had told my wife about them with real confidence, something I had been careful not to do too often, because confidence can curdle into expectation and expectation into crash. But I had let myself feel good about those two. And both came back no.

I did not go for our evening walk that night. I was too heavy. I sat on the couch and could not move much. My wife came and sat beside me. That quiet, deliberate gesture meant more than any words could have. She told me she believed in me. That we would scale through this together. That something would work out.

My son, from across the room, said he had zero doubt I would land something. It was just a matter of time, he said. He said it with the absolute certainty of someone who has never once considered the alternative, and in that moment, his certainty was medicine.

The Worst Morning

That night I could not sleep. I turned the rejections over and over. What had I said wrong? What had I missed? Was I reading the interviews incorrectly, or was there something genuinely broken in how I was presenting myself?

The next morning I woke with a heaviness in my chest that felt almost physical. A real ache, like something pressing from the inside. I had not experienced that before. I sat at my desk and stared at the wall for a while.

The thought that scared me most was not: will I find a job in time? It was: am I still competitive enough for this? Am I cut out for the big leagues anymore?

Nineteen years without an external interview. Eleven at Microsoft. Eight at Nokia before that, where I had been acquired rather than hired from outside. I had built things, filed patents, shipped products used by hundreds of thousands of people. But had I been inside the walls too long? Had I calcified without noticing?

I reached out to the hiring manager at Microsoft AI and asked for feedback. Most companies do not provide it. This one did. The response was mixed: two interviewers had been positive about me, two had not. There were specific areas flagged. It stung. It also clarified. I made notes to prepare better.

I have come to believe something firmly about rejection in hiring: it is almost never a verdict on a person’s worth or capabilities. It is mostly about fit, timing, and the specific chemistry of a moment. That belief did not make the worst morning easier. But it made it survivable.

The Guest WiFi

Months after the layoff, I was speaking at a conference held at the Microsoft office in New York City. I walked into a familiar building, arriving this time as a guest, a speaker, someone brought in from the outside to add value.

When I arrived, my phone reached automatically for the office WiFi network, the way it had done for years. Muscle memory. The phone had connected to Microsoft’s network in buildings across the country and around the world. It knew the routine.

Nothing connected. I had to sign in manually as a visitor. MSFTGUEST.

A small thing. The most mundane possible moment. But those small things have a way of landing hard when you are not expecting them. That login screen told me, quietly and without ceremony, that the relationship was over. The network had moved on. And I had to as well.

The Turn

The turning point came quietly, the way turning points often do. An Amazon recruiter reached out. The pipeline started to slowly build again. There were conversations, then more conversations, then real momentum.

Something shifted inside me. The dread did not disappear. The countdown was still running. But I was back in fight mode.

I changed the number on the post-it. Then I changed it again.

Next: Chapter III, The Rebound. Nineteen years without an external interview, and what it took to become match-ready in weeks.

This is that story.

The Announcement

I had been hearing whispers on Blind for weeks. Something was coming. Still, when the meeting invite landed in my calendar that Tuesday morning in May, labelled simply “Important Business Update,” something in my stomach went cold.

I was at the Microsoft office in Burlington, Massachusetts. A quiet day. Not many people around. I reached out to teammates mostly remote colleagues. Some had received the same invite. I messaged my manager. He had it too. The consensus was : “If you received that invite, it is not good news.”

I had thirty minutes.

There is a particular kind of dread that comes from knowing what is coming and having to sit with it anyway. I think of it now as a sentencing. You know the verdict before the gavel falls. The call itself confirmed everything: the CVP and HR spoke on the line. No cameras allowed. No chat enabled. Just voices, delivering the news of role elimination without warmth, without acknowledgment of who was on the other end. Eleven years, reduced to an organizational adjustment, a line item in a SpreadSheet.

After the call, I spoke with my skip-level, then my manager, then a few long-term colleagues. The office was quiet. I was largely alone with it. The weight of it settled in slowly, the way bad news always does when there is no one to react with you in the same room.

The Last Post

Before I left the office that day, I sat down and wrote. I needed to say something while I still could, while the rawness of the moment was still honest. I opened LinkedIn and typed what was true:

“Your name gets hyphenated with the organization: Ashish Bhatia, Product Manager at Microsoft. And then, an event like today’s suddenly severs you from that anchor point. It is not just a job ending. It feels like losing a part of yourself.”

I ended the post with a line that surprised me as I wrote it:

“It is time for me to de-hyphenate myself.”

What happened next still moves me. Within hours, hundreds of people responded. Colleagues I had built things with. Strangers I had never met. People offering connections, referrals, encouragement, and simply their presence in a moment of grief. The community I had invested in over years showed up immediately. I would later have to hand my phone to my son and ask him to thank people on my behalf and triage for me, who had a concrete lead to share.

But that was later. In that moment, I looked around the office one last time.

On the cubicle wall outside my desk, in my own handwriting, two yellow post-its:

“My Road ends Here.”

“Good bye.”

I grabbed my things and walked out of the building where I had first arrived as a Nokia employee over a decade earlier, before Microsoft’s acquisition. I would not be returning as an employee.

The Drive Home

My hands were shaking on the steering wheel. My mind kept circling the same thought: my stay in this country depends on my job. That job no longer exists. Everything we had built here, our home, our life, the place we called Home, was now in question.

When my wife and son came home from school, I gathered them at the dining table. When I began to speak, they thought someone in the family had died. In a way, something had. I explained what had happened and what it meant for all of us. We held hands around the table.

My son looked at me and said: “Dad, don’t settle for less. You are worth way more than Microsoft ever gave you credit for.”

I promised them we would fight through this together.

The Evening

I told my family I needed to go to a local AI networking event that night. AI Woodstack, hosted by . Part of me had no business being there. The other part understood clearly: networking was the only real path forward, and sitting alone at home with the weight of the day would not help me or my family.

I drove to the event. I told a few close people in the room what had happened that afternoon. Most did not know. I was there, present, moving, connecting. It was the first act of resilience, even if I did not think of it that way at the time. I was simply refusing to stop.

The FedEx Counter

Microsoft asked us to return our equipment. The Friday of that first week, I spent rest of the week finalizing my resume, sending it to friends, responding to the recruiter messages that had already started coming in. I could no longer access the Microsoft building. I was home all day, mostly alone, backing up personal files, immigration documents, a decade of a digital life.

On Friday I drove to a FedEx counter inside a Staples store.

The woman there told me she had never shipped a laptop before. She did not have the right sized box. I found the smallest one that would fit. I wrapped everything carefully in bubble wrap. And then I wrote a note to the admin in Seattle who had supported our team for years:

“Thank you for handling this last shipment from me. Love from Ashish.”

I sealed the box and pasted the shipping label. I handed it back to her and walked out to the parking lot.

My shoulders felt lighter. I cannot fully explain it, and its to comprehend. The badge, the corporate card, the stickers, the machine I had built things on for years. All of it, boxed and shipped and gone. And something in me felt, for the first time since Monday, like I could breathe.

There was a monster of uncertainty ahead of me. But the task was now clear.

Survive or perish.

Next: Chapter II, The 120-Day Clock. When survival has a deadline, and the dark days that followed before the fight came back.

Andrej Karpathy, former Director of AI at Tesla, founding OpenAI member has talked about maintaining a structured markdown wiki that AI agents actively compile and keep up to date. No vector databases, no RAG pipelines. Just disciplined file organization and a capable agent synthesizing everything you feed it.

I built my own version of this in Obsidian. It now covers 120+ clipped articles across topics I care about: GenAI, evaluation frameworks, agentic systems, conversational discovery, responsible AI.

This post is just me sharing what I built so you can replicate it.

How it works

Three folders:

SecondBrain/
├── raw/            ← drop clips here
│   └── processed/  ← clips archived here after ingest
├── wiki/           ← AI-compiled knowledge pages
└── projects/       ← work-specific synthesis

You clip articles. You tell your agent to ingest. It reads each file, maps it to the right wiki pages, writes synthesized knowledge with inline citations back to the source, and moves the files to processed/. The wiki grows over time. Every claim traces back to a source file. Pages link to each other.

The agent does four things: Ingest, Query, Lint, and Compile. That’s it.

The two files

Everything the agent needs to operate the vault lives in two markdown files:

CLAUDE.md — the agent schema. Folder conventions, naming rules, the four operations, knowledge graph linking rules, and enough context about my work so the agent understands what topics matter.

SecondBrain-Setup-Guide.md — the replication guide. The full wiki page template, the ingest workflow step by step, Obsidian plugin list, how to set up scheduled weekly ingestion, and the complete wiki catalog.

You hand these two files to any capable AI agent and say: “Replicate this setup for me.” It builds the structure, initializes the index, and is ready to start ingesting in one session.

Claude.md

About This Vault

This is [Name]’s personal knowledge base, maintained using Andrej Karpathy’s LLM Wiki pattern.
It is a living, AI-compiled wiki of knowledge relevant to Ashish’s work as a Senior Product Manager
in Generative AI, conversational discovery, search & recommendation systems, and product leadership.

[Name]’s Context

• Current role:

• Background:

• Key domains:

• Important topics:

Folder Structure

SecondBrain/
├── raw/              ← Immutable source material. Never edit these files.
│   ├── assets/       ← Locally downloaded images referenced in raw files
│   └── processed/    ← Raw files moved here after ingest (read-only archive)
├── wiki/             ← AI-compiled, structured knowledge pages (YOU write here)
│   └── index.md      ← Master index of all wiki pages with one-line descriptions
├── projects/         ← Project-specific knowledge synthesis pages
│   └── index.md      ← Registry of all active and reference projects
├── _staging/         ← Rough notes or drafts; promote to wiki/ when ready
├── CLAUDE.md         ← This file: schema and instructions for AI agents
└── log.md            ← Chronological log of ingests, queries, and maintenance

Naming Conventions

raw/ files

• Use case: `gpt-4-technical-report.md

• Prefix with source type if useful: article-, paper-, book-, video-, tweet-

• Never rename or edit raw files after ingest

wiki/ files

• Use Title Case with hyphens: Evaluation-Frameworks.md

• One concept per page; avoid catch-all pages

• Every page MUST have YAML frontmatter (see template below)

Wiki Page Template

Every wiki page must start with:

---
title: [Page Title]
tags: [tag1, tag2, tag3]
created: [YYYY-MM-DD]
updated: [YYYY-MM-DD]
sources:
  - "[[raw/processed/filename.md]]"   ← Obsidian wikilink to processed source file
related: [list of other wiki/ pages that link here]
---

Then write the content in clean markdown with:

• A one-paragraph TL;DR summary at the top

• H2 sections for major subtopics

• wikilinks to other pages using Obsidian link syntax

• Inline source citations using [[raw/processed/filename]] immediately after each key claim or insight

• A ## Sources section at the bottom with full wikilinks to every processed source file

Knowledge Graph Linking Rules

These rules ensure the vault is a fully connected knowledge graph — every node links to others.

1. Inline Source Citations (REQUIRED)

After EVERY significant insight, claim, statistic, or concept extracted from a raw file, add an
inline wikilink citation to the processed source:

Insight or claim text here. [[raw/processed/source-filename]]

Example:

AI is reshaping more jobs than it replaces — most roles will change substantially rather than
disappear entirely. raw/processed/AI Will Reshape More Jobs Than It Replaces

This means a reader can jump from any insight directly to the original source in one click.

2. Sources Frontmatter (REQUIRED)

The sources: field in YAML frontmatter must list every processed file that contributed to the page,
as a wikilink:

sources:
  - "[[raw/processed/AI Will Reshape More Jobs Than It Replaces]]"
  - "[[raw/processed/shifting-bottleneck-citations]]"

3. Sources Section (REQUIRED)

Every wiki page must end with a ## Sources section listing all contributing files with wikilinks
AND a one-line description of what each source contributed:

## Sources
- [[raw/processed/AI Will Reshape More Jobs Than It Replaces]] — BCG study on task automation vs job displacement
- [[raw/processed/shifting-bottleneck-citations]] — Evidence that code review is the new AI bottleneck

4. Insight Depth Standard

When extracting insights from raw files into wiki pages:

• Do NOT just copy headings or topic names

• Write a 1–3 sentence explanation of WHAT the insight means and WHY it matters to Ashish’s work

• Include numbers, percentages, or concrete examples where the source provides them

• If a concept is complex, include a sub-section (H3) with full explanation

Bad: > AI affects jobs
Good: `> BCG’s 2026 study found that AI will reshape >70% of tasks in most roles while outright

eliminating fewer than 10% of jobs. The implication for product strategy: build tools that
augment human judgment, not just automate tasks. raw/processed/AI Will Reshape More Jobs Than It Replaces`

5. Cross-Page Wikilinks (REQUIRED)

Every wiki page must link to at least 2 other wiki pages using wikilinks. When you introduce a
concept covered by another page, link to it:

• “...which relates to how Evaluation-Frameworks are used to measure quality...”

• “...a key pattern in Agentic-Web workflows...”

6. wiki/index.md Connectivity

The wiki/index.md master index must:

• Link every wiki page by name using wikilinks

• Include the primary source files (as raw/processed/...) that contributed to each page

• Be updated every time a new page is created or a page gains major new sources

7. projects/ Connectivity

Project pages must link to:

• Relevant wiki pages they draw on (Conversational-Discovery, Evaluation-Frameworks, etc.)

• Any raw source files that are specifically relevant to that project

8. Orphan Prevention

During Lint, flag any wiki page that:

• Has no sources: entries in frontmatter

• Has no inline raw/processed/... citations in the body

• Has no outbound wikilinks to other wiki pages

• Is not referenced by any other page or by wiki/index.md

AI Agent Workflow

Ingest (when new raw/ files are added)

1. Read each new file in raw/ (skip files already in raw/processed/)

2. Identify the core concepts, insights, and claims it covers

3. For each insight, write a clear 1–3 sentence explanation (see Insight Depth Standard above)

4. Check if existing wiki/ pages cover those concepts

   • If YES: append new insights to the relevant section, add inline citations, update sources: frontmatter, update updated: date

   • If NO: create a new wiki/ page using the template, with full frontmatter and Sources section

5. Every insight added to a wiki page MUST include an inline raw/processed/filename citation

6. Update wiki/index.md — add any new pages, and add source file references to existing entries

7. Move the processed raw file to raw/processed/ (create folder if it doesn’t exist)

8. Add an entry to log.md

Query (when asked a question)

1. Read wiki/index.md to identify relevant pages

2. Read the relevant wiki/ pages

3. Answer from compiled knowledge; cite wiki page names AND raw/processed/ sources

4. If the question reveals a knowledge gap, flag it

Lint / Maintenance (weekly or on request)

Scan the entire wiki/ and flag:

• Contradictions: two pages making conflicting claims

• Orphans: pages with no inbound wikilinks from other pages or index

• Uncited insights: claims with no raw/processed/... inline citation

• Stale content: claims referencing outdated sources

• Missing concepts: topics mentioned but not having their own page

• Dead links: wikilinks pointing to non-existent pages

• Unlinked sources: wiki pages whose sources: frontmatter doesn’t link to processed files

Report findings in log.md under a ## Lint [date] heading.

Tone & Style for Wiki Pages

• Write in third-person neutral, encyclopedic style

• Be concise: prefer bullets for lists of items, prose for explanations

• Highlight key terms in bold on first use

• Include concrete examples where possible

• Prioritize Ashish’s domains: GenAI, product strategy, conversational AI, evaluation, responsible AI

• Insight depth over breadth: one well-explained insight is worth more than five shallow bullets

What NOT to Do

• Never delete files from raw/ — they are the source of truth

• Never create wiki pages that duplicate existing ones; merge instead

• Never add speculation without sourcing it to a raw/ file

• Never remove the YAML frontmatter from wiki pages

• Do not create pages for transient or ephemeral information (put that in _staging/)

• Never write a wiki insight without an inline source citation

• Never leave the sources: frontmatter field empty on a page that has been updated via ingest

Ingest Script

A Python ingest script (ingest.py) handles automated wiki updates. To use it:

python3 ~/Projects/SecondBrain/ingest.py

The script:

1. Reads all .md files from raw/ (skipping raw/processed/)

2. Updates the relevant wiki pages with new sections and inline citations

3. Adds each processed filename to the sources: frontmatter of every updated page

4. Moves all processed files to raw/processed/

5. Appends an entry to log.md

Before running, ensure wiki files are saved in UTF-8. If any file was created on macOS with special characters (em dash, smart quotes), convert it first:

import os, glob
vault = os.path.expanduser('~/Projects/SecondBrain')
for fpath in glob.glob(os.path.join(vault, '**/*.md'), recursive=True):
    if 'raw/processed' in fpath:
        continue
    raw = open(fpath, 'rb').read()
    try:
        raw.decode('utf-8')
    except UnicodeDecodeError:
        open(fpath, 'w', encoding='utf-8').write(raw.decode('mac_roman'))
        print('Fixed:', fpath)

SecondBrain-Setup-Guide.md

SecondBrain — AI-Powered Knowledge System

Setup & Replication Guide

Purpose: This document is the single source of truth for replicating Ashish Bhatia’s SecondBrain knowledge system on any machine. It is written to be machine-readable — AI agents (Claude Cowork, OpenClaw, or any markdown-capable LLM) can read this file and execute the setup autonomously.

1. Philosophy — The Karpathy LLM Wiki Pattern

This system is based on a pattern popularised by Andrej Karpathy (former Director of AI at Tesla, founding member of OpenAI). The core insight: instead of vector databases and RAG pipelines, maintain a structured markdown wiki that AI agents actively compile and keep up to date.

Why This Works Better Than RAG

Dimension RAG / Vector DB LLM Wiki Pattern Cost Expensive at scale ~95% cheaper — plain files Reliability Retrieval quality varies Deterministic — AI reads the whole wiki Auditability Black box embeddings Fully human-readable markdown Portability DB lock-in Plain .md files — works anywhere Maintenance Re-embed on changes AI edits files directly Context fit Requires chunking ~400K words fits in modern LLM context window

The Three-Folder Pattern

• raw/ — immutable source material. Web clips, PDFs, articles land here. Never edited after saving.

• wiki/ — AI-compiled, structured knowledge pages. The AI writes and updates these.

• _staging/ — rough drafts and notes; promoted to wiki/ when ready.

2. Compatible AI Agents

This system is designed to work with any AI agent capable of reading markdown files and executing shell commands on the host machine. The two primary agents are:

Claude Cowork

Anthropic’s desktop AI tool. Operates via the Cowork interface with access to the user’s file system, shell, and browser. Supports scheduled tasks natively via the Scheduled section in the sidebar.

Usage: Open a Cowork session. Ask Cowork to build the wiki setup using these instructions. Then instruct Cowork to first look for answer in the local wiki, before issuing a web search to answer an user prompt. Finally, issue natural language commands. Cowork should first search local wiki for answers before going to the web.

OpenClaw

Alternative AI agent compatible with this system. Reads the same CLAUDE.md schema and follows the same ingest/query/lint/compile workflow. Any agent that can read markdown and execute osascript or shell commands on macOS can operate this vault.

Agent-agnostic design: The vault is intentionally not coupled to any specific AI product. CLAUDE.md defines the schema; any compliant agent can execute it. To onboard a new agent, have it read CLAUDE.md first.

Onboarding Any New Agent

When starting a session with a new or unfamiliar agent:

Read the file at ~/Projects/SecondBrain/CLAUDE.md and confirm you understand the vault schema, folder structure, naming conventions, wiki page template, and the four agent operations: Ingest, Query, Lint, and Compile.

3. Prerequisites

Software to Install

Tool Purpose Where Obsidian Markdown vault app — renders the wiki and graph view obsidian.md Obsidian Web Clipper Chrome extension — saves articles to raw/ Chrome Web Store Claude Cowork or OpenClaw AI agent that ingests and compiles claude.ai / OpenClaw docs

No additional dependencies required

Everything is plain markdown files on disk. No databases, no servers, no Python environments, no embeddings pipeline. The vault is a regular folder.

4. Folder Structure

~/Projects/                              ← root projects directory
├── SecondBrain/                         ← Obsidian vault (knowledge layer)
│   ├── raw/                             ← web clips and PDFs land here
│   │   ├── processed/                   ← moved here after ingesting
│   │   └── assets/                      ← locally saved images
│   ├── wiki/                            ← AI-compiled knowledge pages
│   │   └── index.md                     ← master page catalog
│   ├── projects/                        ← knowledge synthesis per project
│   │   └── index.md                     ← project registry
│   ├── _staging/                        ← rough drafts before promotion
│   ├── CLAUDE.md                        ← AI agent schema & instructions
│   └── log.md                           ← chronological ingest history
└── [your-other-projects]/               ← code + artifacts (builder agents work here)

Critical separation: ~/Projects/[project-name]/ contains actual code, configs, and artifacts. SecondBrain/projects/ contains only lightweight knowledge synthesis markdown — goals, decisions, metrics, open questions. Builder agents (Cursor, Claude Code, OpenClaw in coding mode) work in the project folder and never need to touch SecondBrain.

5. Step-by-Step Setup

Step 1 — Create the folder structure

Run the following in Terminal:

mkdir -p ~/Projects/SecondBrain/raw/processed
mkdir -p ~/Projects/SecondBrain/raw/assets
mkdir -p ~/Projects/SecondBrain/wiki
mkdir -p ~/Projects/SecondBrain/projects
mkdir -p ~/Projects/SecondBrain/_staging

Step 2 — Create CLAUDE.md

CLAUDE.md is the brain of the system — it tells AI agents how to maintain the wiki. Create it at ~/Projects/SecondBrain/CLAUDE.md.

Required sections:

• About This Vault — who owns it, what domains it covers

• Owner Context — current role, background, key domains, active projects

• Folder Structure — annotated diagram

• Naming Conventions — kebab-case for raw/, Title-Case-With-Hyphens for wiki/

• Wiki Page Template — YAML frontmatter schema (see Section 6 below)

• AI Agent Workflow — Ingest, Query, Lint, Compile operations with step-by-step instructions

• Tone & Style — encyclopedic, third-person, concise, domain-aware

• Prohibitions — never delete raw/, never duplicate pages, never speculate without sources

Tip: Copy the existing CLAUDE.md verbatim to the new machine — it contains all agent instructions calibrated to Ashish’s domains.

Step 3 — Create log.md and wiki/index.md

log.md tracks every ingest, query, and maintenance operation chronologically.

## YYYY-MM-DD — [Operation type]

**Operation**: Ingest / Query / Lint / Compile  
**Files processed**: [list]  
**Wiki pages updated/created**: [list]  
**Notes**: [summary]

wiki/index.md is the master catalog — every wiki page gets a one-line entry in a table organised by category.

Step 4 — Open the vault in Obsidian

1. Open Obsidian

2. Click Open folder as vault from the vault picker

3. Navigate to ~/Projects/SecondBrain/ and click Open

4. Obsidian indexes all markdown files and renders [[wikilinks]] automatically

Note: If the vault shows as “not found” after moving or renaming the folder, always use “Open folder as vault” rather than the obsidian:// deep link — the deep link relies on a cached path that becomes stale after moves.

Step 5 — Install and configure Obsidian Web Clipper

1. Search Obsidian Web Clipper in the Chrome Web Store and install

2. Click the extension icon → go to Settings

3. Under General → set Vault to SecondBrain

4. Under Templates → select Default template → set Note location to raw

5. Confirm the Note content field contains {{content}} — this captures the article body

Default template properties:

title:       {{title}}
source:      {{url}}
author:      {{author|split:", "|wikilink|join}}
published:   {{published}}
created:     {{date}}
description: {{description}}
tags:        clippings

Known Web Clipper limitations:

• PDFs in Chrome’s PDF viewer cannot be clipped — the extension can’t access the sandboxed renderer

• Heavily JS-rendered pages may only clip metadata, not body content (e.g. some consulting firm sites)

• Workaround: paste the URL into your AI agent session and say: "Fetch this URL and write the content to raw/: [url]"

Step 6 — Seed the wiki

Create the initial wiki pages covering your primary domains. See Section 7 for the full list. Each page must follow the YAML frontmatter template in Section 6.

Step 7 — Set up the automated weekly ingest

Ask your AI agent (Claude Cowork or OpenClaw):

Set up a scheduled task to ingest raw/ files in my SecondBrain vault every Monday at 8am.
The vault is at ~/Projects/SecondBrain/. The agent should read CLAUDE.md first, then process
all unprocessed files in raw/, update wiki pages, move processed files to raw/processed/,
and append a log entry to log.md.

Important: After creating the task, run it once manually to pre-approve tool permissions — this prevents future automated runs from pausing on approval prompts.

6. Wiki Page Template

Every wiki page must begin with this YAML frontmatter:

---
title: [Page Title]
tags: [tag1, tag2, tag3]
created: YYYY-MM-DD
updated: YYYY-MM-DD
sources:
  - "[[raw/processed/source-filename]]"   # Obsidian wikilink to the processed source file
related: []  # list of other wiki/ pages that link here
---

Changed in v1.2: sources: entries must now be full Obsidian [[wikilinks]] to raw/processed/ files, not plain filenames. This connects the knowledge graph so you can click from a wiki page directly to its source clip.

Body structure:

# Page Title

One-paragraph TL;DR summary at the top.

## Subtopic 1

Content with **bold** key terms on first use. Use [[wikilinks]] to other pages.
Every significant insight, claim, or statistic must be followed immediately by an
inline source citation: `[[raw/processed/source-filename]]`

Example:
> BCG found AI will reshape >70% of job tasks while eliminating fewer than 10%
> of roles outright. [[raw/processed/AI Will Reshape More Jobs Than It Replaces]]

## Subtopic 2

...

## Sources

- [[raw/processed/source-filename]] — one-line description of what it contributed
- [[raw/processed/another-source]] — what this source added

Three mandatory citation layers (all required on every wiki page):

1. sources: frontmatter — [[raw/processed/...]] wikilinks to every source file

2. Inline citations — [[raw/processed/...]] immediately after every insight in the body

3. ## Sources footer — wikilinks + one-line description of each source’s contribution

7. Wiki Pages — Full Catalog

Core Domains ([Name] expertise)

[Your Domains]

Areas of Active Interest

[Your areas of interest]

Meta

Page file Category Description LLM-Wiki-Pattern.md Meta The Karpathy pattern — how this system itself works

8. Project Knowledge Pages

The projects/ folder holds a lightweight knowledge synthesis layer for active work. It is not a duplicate of your project folder — it captures the thinking layer: goals, open decisions, success metrics, and links to wiki pages.

projects/index.md

A registry table of all projects — active and completed. Format:

| Project | Role | Status | Related Wiki |
|---------|------|--------|-------------|
| [[Project-Name]] | Sr. PM | 🟢 Active | [[Wiki-Page-1]], [[Wiki-Page-2]] |

Per-project page structure

---
title: [Project Name]
project_status: active | completed | paused
related: [wiki pages]
---

## Project Context
## Problem Space
## Key Goals
## Stakeholders & Partners
## Technical Approach
## Success Metrics (table)
## Key Decisions & Open Questions (checkbox list)
## Related Wiki Pages ([[wikilinks]])

Rule: when a project ends, archive its page with a final summary. The wiki pages it linked to persist — they accumulate knowledge across projects over time.

9. Ingest Workflow

Manual ingest

Open an AI agent session and issue one of these prompts:

"Ingest the new articles in raw/"
"What did I clip this week about conversational AI?"
"Add anything relevant from raw/ to my Audible project page"
"Fetch this URL and save it to raw/: https://..."

What the agent does during ingest

1. Reads CLAUDE.md for schema and conventions

2. Lists files in raw/ — skips processed/ and assets/

3. For each file: reads content → maps to wiki page(s) → appends synthesis section

4. Every extracted insight must include an inline [[raw/processed/filename]] citation — see Knowledge Graph Rules in CLAUDE.md

5. Adds [[raw/processed/filename]] to the sources: frontmatter of every updated page

6. Creates new wiki pages if no existing page matches the topic

7. Adds relevant notes to active project pages in projects/

8. Moves processed files to raw/processed/ — never deleted

9. Updates wiki/index.md if new pages were created, adding source links per the Knowledge Graph Rules

10. Appends a structured entry to log.md

Handling empty clips

If a clipped file contains only frontmatter (no body content), the agent should:

• Note it in the log as “empty clip — skipped”

• Move it to raw/processed/ anyway

• Optionally attempt to fetch the URL from the source: frontmatter field

• Add a stub entry in AI-Papers.md (or relevant wiki page) with a re-clip reminder

Files requiring re-clip (as of 2026-04-14):
| File | Re-clip source |
|------|---------------|
| AI Will Reshape More Jobs Than It Replaces.md | BCG article — use Reader Mode |
| Forecasting the Economic Effects of AI.md | PDF — check source URL in raw file |
| SkillNet - Create, Evaluate, and Connect AI Skills.md | https://arxiv.org/pdf/2603.04448 |
| Meta-Harness - End-to-End Optimization of Model Harnesses.md | https://arxiv.org/pdf/2603.28052 |

Handling encoding errors

macOS can produce files with Mac Roman encoding instead of UTF-8 (em dashes, smart quotes).
If the ingest script raises UnicodeDecodeError, run this one-time fix before re-running:

import os, glob
vault = os.path.expanduser('~/Projects/SecondBrain')
for fpath in glob.glob(os.path.join(vault, '**/*.md'), recursive=True):
    if 'raw/processed' in fpath:
        continue
    raw = open(fpath, 'rb').read()
    try:
        raw.decode('utf-8')
    except UnicodeDecodeError:
        open(fpath, 'w', encoding='utf-8').write(raw.decode('mac_roman'))
        print('Fixed:', os.path.basename(fpath))

Automated weekly ingest

A scheduled task runs every Monday at 8:00 AM. On macOS with Claude Cowork, the task is stored at:

~/Documents/Claude/Scheduled/secondbrain-weekly-ingest/SKILL.md

To replicate on a new machine: open your AI agent (Claude Cowork or OpenClaw) and ask it to create a weekly Monday 8am ingest task for the vault at ~/Projects/SecondBrain/. Provide the full ingest instructions above as context.

10. Day-to-Day Usage

Clipping articles

• Navigate to any article in Chrome

• Click the Web Clipper extension icon

• Confirm vault = SecondBrain, folder = raw, then clip

• For PDFs or JS-rendered pages: paste the URL into your agent and say “Fetch this and write it to raw/”

Querying your knowledge base

Ask your agent naturally — it reads CLAUDE.md and the wiki to answer:

"What do I know about multi-turn conversation state management?"
"Summarise my notes on evaluation frameworks for GenAI"
"What open questions do I have on the Audible project?"
"What AI papers have I collected on memory and retrieval?"
"What's the difference between LangGraph and AutoGen?"

Adding new wiki topics

Either clip an article and let the weekly ingest create the page, or ask directly:

"Add a wiki page for [topic] — here's why I'm interested: [context]"

Weekly review

Every Monday morning, the scheduled task processes the week’s clips and updates the wiki. Review log.md to see what was ingested and what new pages were created.

11. Replication Checklist

Use this checklist when setting up the system on a new machine. An AI agent can execute all steps autonomously if given this document.

# Step Command / Action Done 1 Install Obsidian obsidian.md — do not enable iCloud sync ☐ 2 Create folder structure mkdir -p commands in Section 5, Step 1 ☐ 3 Copy CLAUDE.md From ~/Projects/SecondBrain/CLAUDE.md on old machine ☐ 4 Copy wiki/ folder All .md files — your compiled knowledge base ☐ 5 Copy projects/ folder Project knowledge synthesis pages ☐ 6 Copy log.md Ingest history — optional but useful ☐ 7 Open vault in Obsidian Open folder as vault → ~/Projects/SecondBrain/ ☐ 8 Install Web Clipper Chrome Web Store → Obsidian Web Clipper ☐ 9 Configure Web Clipper Vault = SecondBrain, Note location = raw, body = {{content}} ☐ 10 Set up scheduled ingest Ask agent to create weekly Monday 8am ingest task ☐ 11 Run ingest once manually Pre-approve tool permissions for future automated runs ☐ 12 Verify wiki/index.md Confirm all pages are listed correctly ☐

Minimum viable migration: if you only copy three things, copy CLAUDE.md, wiki/, and projects/. Everything else can be recreated from scratch.

12. Quick Reference

Key file paths

File Purpose ~/Projects/SecondBrain/CLAUDE.md AI agent instructions — read this first every session ~/Projects/SecondBrain/wiki/index.md Master catalog of all wiki pages ~/Projects/SecondBrain/projects/index.md Registry of all projects ~/Projects/SecondBrain/log.md Chronological ingest history ~/Projects/SecondBrain/raw/ Drop clips here ~/Projects/SecondBrain/raw/processed/ Clips moved here after ingest ~/Documents/Claude/Scheduled/secondbrain-weekly-ingest/ Scheduled task (Claude Cowork)

Useful agent prompts

"Ingest the new articles in raw/"
"What do I know about [topic]?"
"Add a wiki page for [topic]"
"Lint the wiki — find orphans and contradictions"
"Summarise open questions across all my project pages"
"Fetch this URL and save it to raw/: [url]"
"What new wiki pages were created this month?"
"Update my Audible project page with anything relevant from this week's clips"
"Add inline citations to [wiki page] — link each insight to its source in raw/processed/"
"Lint the wiki for orphan pages and missing source citations"
"Add a Sources section to all wiki pages that are missing one"

Agent onboarding prompt (for any new AI agent)

Read ~/Projects/SecondBrain/CLAUDE.md and ~/Projects/SecondBrain/SecondBrain-Setup-Guide.md.
Confirm you understand: the vault schema, folder structure, naming conventions, wiki page
YAML frontmatter template, and the four operations: Ingest, Query, Lint, and Compile.
Then list the current wiki pages from wiki/index.md.

13. Knowledge Graph Rules (Summary)

Full rules are in CLAUDE.md under Knowledge Graph Linking Rules. This is the quick-reference version.

The vault is designed as a fully connected knowledge graph — every wiki page links to its sources, and every source can be traced back to the original clip. There are eight rules:

1. Inline citations — after every significant insight: [[raw/processed/source]]

2. Sources frontmatter — every source file as a [[raw/processed/...]] wikilink in YAML

3. Sources footer — ## Sources section at the bottom of every page with wikilinks + descriptions

4. Insight depth — 1–3 sentences explaining what the insight means and why it matters; include numbers

5. Cross-page links — every page links to at least 2 other wiki pages via [[wikilinks]]

6. index.md connectivity — wiki/index.md lists every page and its primary source files

7. projects/ links — project pages link to both wiki pages and relevant raw source files

8. Orphan prevention — during Lint, flag pages with no sources, no inline citations, or no inbound links

Current connectivity gaps (as of 2026-04-14):

• wiki/index.md — needs [[raw/processed/...]] source links added to existing entries

• All wiki pages — ## Sources footer section not yet added to pre-April-2026 pages

• Older pages — inline citations only exist on content added after April 2026

This document is part of the SecondBrain vault and should be kept at ~/Projects/SecondBrain/SecondBrain-Setup-Guide.md so it is accessible to all AI agents operating on this vault.

When generative AI started making its way into enterprise workflows, the industry rallied around a reassuring narrative: humans must stay in the loop. It felt responsible. It felt safe. And for a while, it was useful.

But here’s what I’ve come to believe: human-in-the-loop was always a temporary narrative. A necessary one at the time to manage the transition, to calm the outcry, to buy time for adoption. The long-term arc was always toward human-optional or fully automated execution. Not because humans don’t matter. But because anytime you put a human in the middle of a process, you’ve created a bottleneck. And every system, eventually, wants to eliminate its bottlenecks.

The Bottleneck Always Shifts Upstream

Here’s what actually happens when organizations adopt AI at scale.

You automate the high-volume, repetitive work. Velocity spikes. You celebrate. Then you realize you’ve solved one problem and created another because the constraint didn’t disappear. It moved. It shifted to the next layer up, usually to your most expensive, most experienced people.

That pattern is playing out right now across industries. Here are three examples.

Example 1: Code Review is the New Bottleneck

When coding agents like Claude Code entered the picture, development teams saw immediate productivity gains. Developers could ship more code and at a faster clip. What could go wrong?

The PR queue starts to build up.

Cognition, the company behind Devin, said it plainly: “We’re hearing from our customers that code review — not code generation — is now the bottleneck to shipping great products.” (source) They even built a separate product, Devin Review specifically because of this problem.

The New Stack reported the same dynamic: “Using AI to increase velocity means that more code is being thrown over the wall, and someone has to review it. We’re actually starting to see that code review is becoming the new bottleneck.” (source)

Your principal engineers and senior developers — your most expensive, most strategic humans — are drowning in pull requests. They can’t keep up. So what do you do? You automate review too. But then who’s accountable when something slips?

The bottleneck shifted. And now it’s moving again.

Example 2: Radiology — Faster Scans, Harder Decisions

Radiology is one of the most advanced examples of AI entering a high-skill profession. AI can now flag anomalies in imaging studies before a radiologist even opens the file. It automates quality assurance, eliminates routine screening work, and dramatically increases throughput.

And yet, demand for radiologists is actually growing, up 9% even as AI adoption accelerates. (source)

Why? Because AI handled the volume. What’s left are the hard cases. The anomalies that don’t fit standard patterns. The clinical context that requires judgment. The communication of nuanced findings to a patient’s care team.

Nature published research on this directly, AI tools reduce routine workload, but radiologists get increasingly concentrated on edge cases and high-stakes interpretation. (source)

The bottleneck didn’t disappear. It moved upstream to the most complex, highest-consequence decisions in the workflow. And the human sitting at that chokepoint is now under more pressure than ever.

Example 3: Recruiting — The Screener Isn’t the Problem Anymore

Hiring at scale used to break down at the top of the funnel. Thousands of resumes, limited recruiter time, qualified candidates slipping through. AI fixed that.

Unilever used AI to filter out 80% of applicants, cutting time-to-hire by 90% and saving over 50,000 hours of recruiter time. (source)

But here’s what happens next. The hiring manager’s calendar fills up. The candidates AI passed still need to be interviewed. Culture fit, leadership potential, team dynamics, those aren’t things you automate away. A survey by Insight Global found 93% of hiring managers believe human involvement remains essential even with AI screening in place. (source)

The bottleneck didn’t go away. It moved from resume screening to final-stage evaluation, which by the way, is exactly where your senior leadership spends time.

The Pattern is the Point

Code review. Radiology. Recruiting. Three different industries, same story.

You automate the volume layer. Throughput increases. Then the constraint migrates to wherever human judgment is hardest to replace. And the people sitting at that new chokepoint are your most valuable, most expensive resources; principal engineers, senior radiologists, hiring executives.

This is the trap that most organizations are walking into right now. They measure AI adoption by deployment. “We gave everyone a coding agent”. “We launched an AI screener”. They celebrate the initial productivity gains and stop there. They don’t realize they’ve shifted the problem, not solved it.

Where Humans Actually Belong

Here’s the framework I keep coming back to.

There are two ends of any workflow that should always be in human hands:

Specification — What needs to be done? What problem are we solving? What does success look like? This is intent-setting. It’s strategic. It cannot and should not be delegated.

Validation — Did it work? Were the objectives actually met? Did the outcome match the intent? This is verification. It requires judgment. It also cannot and should not be delegated just yet.

Everything in between is implementation detail. Writing the code. Screening the applicant. Flagging the scan. That is where machines should operate. That is where you keep pushing automation.

But here’s the critical point: you can only hold those two ends of the thread if the middle is clear. If humans are stuck somewhere in the implementation layer; reviewing code, screening edge cases, manually processing what automation missed, then they can’t focus on specification and validation. The whole system degrades.

The Job of Organizational Leaders

If you’re a VP, a product leader, or a decision-maker in an AI-forward organization, this is what I want you to take away:

AI adoption is not a finish line, that was last year’s goal. Identify the human bottlenecks.

Every time you automate a layer, the constraint moves upstream. Your job is to see it coming, name it, and eliminate it, before it becomes the new chokepoint that slows everything down.

The cycle looks like this:

1. Automate the current bottleneck

2. Watch where the constraint migrates

3. Identify the new chokepoint

4. Plan to eliminate it

5. Repeat

Humans are not going away. But they will keep moving to higher and higher layers of judgment, until either those layers get automated too, or we reach a decision of such consequence that we decide the liability simply must sit with a person.

Until that moment, the job is to keep clearing the path. Eliminate the bottleneck. Hold the two ends of the thread. Let machines handle everything in between.

That’s how you lead in an AI-first world.

Sources

1. Cognition (Devin) — Code Review as the New Bottleneck

   • Source: Cognition Blog — “Devin Review: AI to Stop Slop”

   • URL: https://cognition.ai/blog/devin-review

2. Cognition Blog — Devin Annual Performance Review

   • URL: https://cognition.ai/blog/devin-annual-performance-review-2025

3. Devin Docs — Devin Review

   • URL: https://docs.devin.ai/work-with-devin/devin-review

4. The New Stack — Is AI Creating a New Code Review Bottleneck for Senior Engineers?

   • URL: https://thenewstack.io/is-ai-creating-a-new-code-review-bottleneck-for-senior-engineers/

5. UDS Health — AI in Radiology: Why Demand for Humans is Growing 9%

   • URL: https://udshealth.com/blog/ai-radiology-demand-for-humans-growing/

6. Nature — AI Solutions to the Radiology Workforce Shortage

   • URL: https://www.nature.com/articles/s44401-025-00023-6

7. Oxford Academic — BJR Artificial Intelligence: Radiology Report Generation

   • URL: https://academic.oup.com/bjrai/article/3/1/ubag003/8445887

8. HeroHunt.ai — AI-Driven Candidate Screening: The 2025 In-Depth Guide

   • URL: https://www.herohunt.ai/blog/ai-driven-candidate-screening-the-2025-in-depth-guide/

9. x0pa.com — AI Resume Screening

   • URL: https://x0pa.com/glossary/ai-resume-screening/

To put that in AI terms: global AI compute capacity across all facilities worldwide currently sits at approximately 20 gigawatts. Terafab's one-terawatt target is 50 times that figure, surpassing every semiconductor fab on the planet combined. To put it in perspective: the Hoover Dam produces about two gigawatts. One terawatt equals 500 Hoover Dams. That's not hyperbole. That's the infrastructure play Musk, Tesla, SpaceX, and xAI are building together, with Intel now joining as a foundry partner at a pilot cost of $20–25 billion.

This matters because compute is about to become what oil was for the twentieth century. And like oil, it will not be distributed equally.

The Energy Floor

Let's start with the constraint everyone's aware: energy.

China generates roughly 10,000 terawatt-hours of electricity annually. The US produces around 4,635 TWh. Japan, about 1,000 TWh. Russia, 1,209 TWh. Global electricity production sits around 28,000–30,000 TWh per year.

One terawatt of compute capacity. That's 3–4% of everything the world produces in electricity. Every year. Added on top.

But here's what's often missed: countries don't dedicate all their electricity to compute. Energy powers agriculture, transportation, manufacturing, homes. The question isn't how much electricity you produce. It's how much spare electricity you can afford to dedicate to AI infrastructure.

That spare capacity is concentrated. The US is building 8.2 gigawatts of AI data center capacity today. China has 4.5 gigawatts in mainland facilities. Japan, 1.5 GW. India, 1.3 GW. Australia, 1.3 GW. Europe consolidated — Germany, France, and the EU block — sits around 3 GW total. The UK separately, roughly 0.5 GW.

The gap is stark. And it's widening fast.

The Strategic Allocation Question

Energy capacity alone doesn't tell the full story. You also need strategic intent.

Some countries are all-in on compute infrastructure. China's intelligent computing capacity reached 725.3 EFLOPS in 2024, up 74.1% year-over-year. The US is pushing hard with 2,528 live data centers in 2026, projected to grow to 3,226 by 2031. A few others are making serious moves.

But many nations haven't made bets at the same level. They're not allocating capital, not building data centers, not treating compute as a strategic asset. Some are outright oblivious to what's happening.

The Global South. Most of Africa. Parts of Southeast Asia. Their governments aren't thinking about compute infrastructure the way US and Chinese policymakers are. And it means the gap compounds. Every year they don't build, the leaders extend their lead.

The Intelligence Layer

Now add a second dimension: who's building the frontier models?

Open source has democratized AI in important ways. DeepSeek V3.2 delivers roughly 90% of frontier performance at 1/50th the cost. Qwen 3.5 ships under Apache 2.0 free, commercially permissive, downloadable. Llama 4 runs on your own hardware.

But there's a gap between what's open and what's frontier. That gap matters.

Claude, GPT-5.4 Gemini 3.1 Pro, these are built by US companies. DeepSeek and Qwen come from China. The rest of the world is working with what trickles down to them, yesterday's models, open-source approximations, licensed access.

And the companies controlling frontier development also control distribution. Anthropic decides which nations get access to Claude. OpenAI controls access to GPT models. China controls what leaves its borders.

That's not a technical problem. That's a power play.

Consider what's happening right now. In April 2026, Anthropic signed a single agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, coming online in 2027 — with the vast majority sited in the United States. One company. One deal. Securing compute that rivals the entire AI data center footprint of India, Japan, or Australia. The companies building frontier models aren't just controlling access to intelligence. They're actively cornering the infrastructure required to produce the next generation of it.

Key insight: The price gap between frontier closed models and open alternatives is as wide as 25x. The US controls closed-frontier access. China competes on cost and open weights. Everyone else picks from what's made available to them.

The Outcome

Here's where this leads.

Countries with both frontier compute capacity and frontier model development, the US and China — will set the rules. They'll control which intelligence gets deployed where. They'll use frontier AI as a strategic lever in defense, finance, economics, and any sector that matters.

Intelligence will get nationalized in some form. Not necessarily formally. But governments will have significant say in what potency of AI gets distributed broadly and what gets reserved for strategic use. The decisions being made today, what to build, where to build it, who gets access are geopolitical decisions dressed up as technology decisions.

Everyone else will compete for access to yesterday's models. Hope the open source models are good enough. Watch their competitive position erode as the gap widens.

The countries that understand this now and build accordingly will lead. The ones that don't will follow. And some won't even realize the race started until it's over.

Compute is the new oil. The reserves are being claimed. The question is whether your country is in the room.

Sources

1. Terafab announcement and Intel partnership — Reuters, April 7 2026

2. Terafab scale and pilot cost — BusinessToday India, April 8 2026

3. Terafab announcement location and joint venture details — how2shout

4. Global electricity production rankings — Visual Capitalist

5. Country electricity production figures — Statista

6. Global electricity total — Statista

7. US AI data center capacity — InfotechLead

8. Asia-Pacific AI data center capacity by country — Brightlio / Cushman & Wakefield

9. China intelligent computing capacity — China Economic Net / Global Computing Power Trends Report 2026

10. US data center count and growth projections — InfotechLead

11. Frontier AI model pricing and benchmarks — morphllm.com

12. DeepSeek vs frontier cost comparison — buildfastwithai.com

13. Open source model landscape — Medium / Sanjeev Patel

14. China AI market share growth — Medium / Marc Bara

15. OECD framework on national compute capacity — OECD.AI

16. Global AI compute concentration — sanchez.vc GeoCoded Report

17. India electricity growth — globalelectricity.org

18. Anthropic–Google–Broadcom compute partnership — Anthropic, April 2026

That single finding represented nearly a fifth of all high-severity Firefox vulnerabilities remediated in all of 2025. And here is the part that stopped me cold: the first bug was found within 20 minutes of starting. This wasn't Anthropic's most powerful model. This was Claude Opus 4.6, a model that is already publicly available. We are not ready for what comes next.

Meet Mythos

Last couple of weeks of Anthropic leaks brought Claude Mythos into public consciousness. It is the most capable AI model they have ever built, and they have decided not to release it to the public.

In the this cut throat race to build super intelligence, a company whose entire business is built on releasing AI models has built one so capable that they are deliberately keeping it out of public hands.

Some are calling it the "God Model," and it is easy to see why the label gets attached. Mythos represents a capability threshold that the industry has been nervously anticipating. It is the point where AI stops being a tool that assists security researchers and starts being something that can outperform nearly all of them.

The benchmarks back this up. On CyberGym, which tests real vulnerability reproduction in live open-source software, Mythos Preview scores 83.1%. The previous best, Claude Opus 4.6, scores 66.6%. On SWE-bench Verified, a broad coding benchmark, Mythos hits 93.9% versus Opus 4.6's 80.8%. Does this model belong to a different league of its own, “God Model” perhaps.

The Firefox comparison makes it visceral. When researchers ran Mythos modelsagainst the same Firefox exploit challenge, Mythos produced 181 successful exploits.

What It Has Already Found

To understand what this model is capable of, you have to look at what it has already done in just a few weeks of limited, controlled use.

• Mythos discovered a 27-year-old bug in OpenBSD, an operating system known specifically for its security record. The flaw had survived nearly three decades of human review, audits, and automated scanning. The model found it, and it was critical.

• It found a 16-year-old vulnerability in FFmpeg, the multimedia framework that sits underneath a significant portion of the world's video infrastructure. The bug had survived more than five million automated tool interactions over its lifetime without detection.

• It uncovered a 17-year-old remote code execution vulnerability in FreeBSD (CVE-2026-4747) that allowed unauthenticated root access. The kind of vulnerability that, in the wrong hands, would be worth an enormous amount to a nation-state attacker.

• It also autonomously chained multiple Linux kernel vulnerabilities together, and constructed complex exploits that escaped both browser renderer sandboxes and operating system sandboxes without human guidance.

• Across all of this, Anthropic researchers found "thousands" of additional high and critical severity vulnerabilities spanning every major operating system and every major web browser.

The word "thousands" should not pass by quietly. This is infrastructure we depend on everyday, to manage bank account, pay bills, do office work, etc. These our cornerstones of our digital life.

Project Glasswing: The Defense

Anthropic knows exactly what they have built. Their own internal assessment described Mythos as "currently far ahead of any other AI model in cyber capabilities" and warned it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."

Project Glasswing is their response to that reality.

Rather than shelving the model, Anthropic has structured a controlled, time-limited program that makes Mythos Preview available exclusively to a coalition of organizations responsible for the world's most critical software infrastructure. The goal is to use Mythos offensively, in a controlled way, to find and fix vulnerabilities before bad actors can use a similar model to find and exploit them.

The founding coalition of twelve organizations covers every critical layer of the internet's architecture, and that is not an accident.

• Cloud and Hyperscale Infrastructure (AWS, Google, Microsoft) form the backbone of global compute. Virtually every application, service, and platform runs on or through these three providers. A vulnerability at this layer does not affect one company. It affects everything built on top of it.

• Pure-Play Cybersecurity (CrowdStrike, Palo Alto Networks) brings threat intelligence and security research depth. They understand the attacker's perspective and have the tooling and telemetry to operationalize what Mythos finds.

• AI and Compute (Anthropic, NVIDIA) represent the model itself and the silicon it runs on. NVIDIA's CUDA stack underpins virtually all AI workloads and is itself a high-value target. Having the infrastructure layer inside this initiative is important.

• Networking and Semiconductor Infrastructure (Cisco, Broadcom) own the physical and logical plumbing of the internet. Cisco controls a dominant share of enterprise networking hardware. Broadcom, after acquiring VMware, is foundational to enterprise compute globally.

• Consumer and Device Ecosystems (Apple, Google) together control iOS and Android, the two operating systems running on billions of end-user devices. Chrome is the world's most-used browser. A vulnerability at this layer is not a business problem. It is a personal one, for billions of people.

• Financial Services (JPMorganChase) represents a sector that is both among the highest-value attack targets on earth and an operator of massive, proprietary critical infrastructure. Their inclusion signals that this initiative is not just about software. It is about the systems economies run on.

• Open Source Governance (Linux Foundation) acts as the bridge between the twelve founding members and a wider group of over 40 additional open-source projects and infrastructure maintainers who also have access to Mythos under this program.

Anthropic is committing $100 million in model usage credits to this effort, $2.5 million to Alpha-Omega and OpenSSF, and $1.5 million to the Apache Software Foundation. Findings will be shared publicly within 90 days.

The Problem Glasswing Cannot Solve

Here is what I keep thinking about. Glasswing is starting at right place. The organizations in this coalition represent the most fundamental building blocks of the internet. If you had to choose where to start, this is where you start.

But the internet is not just its building blocks.

Sitting on top of these foundations is an enormous long tail: hundreds of thousands of SaaS products, fintech applications, healthcare platforms, e-commerce systems, APIs, and startup codebases that have no Glasswing access, no Mythos partnership, and no 90-day head start. When a Mythos-class model eventually becomes commercially available, or gets replicated by a well-resourced actor, that entire long tail will be exposed simultaneously.

Glasswing is not a complete solution. It is a head start for the most critical layer. What happens beneath that layer is a much harder problem, and the clock is already running.

What This Means for Us

Most people will read about Project Glasswing and feel a comfortable distance from it. Vulnerabilities in operating systems and browser engines feel abstract. They are not.

Your banking app runs on infrastructure with vulnerabilities. Your healthcare provider's portal runs on frameworks with vulnerabilities. The authentication system that protects your email almost certainly depends on open-source libraries that have never been scanned by anything close to Mythos.

The security researchers and penetration testers who currently protect these systems are talented, but they are working at human speed. Mythos operates at a different scale entirely. Penetration testers estimated that certain exploits the model developed overnight would take their best people several weeks of manual work. That gap in speed and economics is the real disruption. The threat surface is expanding in ways that human-only security teams are not equipped to match.

Anthropic's decision to build Glasswing before releasing Mythos is the right call. Although, this is not a permanent fix. It is a head start. The gap between what AI can do and what most people believe AI can do has never been wider. And that gap is closing from one direction only.

So here is the closing thought I want to leave you with.

A 27-year-old vulnerability in one of the most security-conscious operating systems ever built. Found by AI. Cost of the run: under $50.

That is not a future scenario. That is last month. The ground under all of us is shifting, and the only honest response is to lean in, understand what these models can actually do, and start building new mental models for what comes next.

Sources

1. Project Glasswing: Securing critical software for the AI era — https://www.anthropic.com/glasswing

2. Claude Mythos Preview — https://red.anthropic.com/2026/mythos-preview/

3. Partnering with Mozilla to improve Firefox's security — https://www.anthropic.com/news/mozilla-firefox-security

4. Anthropic's Claude found 22 vulnerabilities in Firefox over two weeks — https://techcrunch.com/2026/03/06/anthropics-claude-found-22-vulnerabilities-in-firefox-over-two-weeks/

5. Anthropic's Claude uncovers 22 Firefox security vulnerabilities — https://www.axios.com/2026/03/06/anthropic-mozilla-claude-opus-bug-hunting

6. Anthropic is giving some firms early access to Claude Mythos — https://fortune.com/2026/04/07/anthropic-claude-mythos-model-project-glasswing-cybersecurity/

7. Anthropic says its most powerful AI cyber model is too dangerous to release publicly — https://venturebeat.com/technology/anthropic-says-its-most-powerful-ai-cyber-model-is-too-dangerous-to-release

8. Anthropic limits Mythos AI rollout over fears hackers could use model for cyberattacks — https://www.cnbc.com/2026/04/07/anthropic-claude-mythos-ai-hackers-cyberattacks.html

9. Introducing Project Glasswing — https://www.linuxfoundation.org/blog/project-glasswing-gives-maintainers-advanced-ai-to-secure-open-source

But Oracle isn’t the headline. Oracle is just the latest data point in something much larger.

Since January 2025, the tech industry has cut over 330,000 jobs. Microsoft: ~19,000. Amazon: ~30,000. Intel: ~34,000. Meta: 3,600. Block: 4,000 - nearly half its workforce. And now Oracle: 30,000. In 2026 alone, we are already at 85,000+, at a pace of nearly 1,000 people per day. AI was explicitly cited as the reason in nearly 1 in 4 layoff announcements this year.

Let’s be honest about what’s actually happening here.

This is not AI replacing jobs. This is capital substitution of labor. Companies are taking on enormous leverage against an AI future that hasn’t paid off yet, and immediately cutting workforce to service that debt. When you raise $25 billion in bonds as part of a $45 billion capital raise, the interest clock starts on day one. AI infrastructure returns take years. Workforce costs can be cut today. TD Cowen did the math in January: eliminate 20-30K people, free up $8-10 billion in annual cash flow. Make the bond payments. Hold the stock price.

Thirty thousand people paid for that calculation with a cold email before sunrise.

What this accounting misses is everything that doesn’t appear on a balance sheet. The institutional knowledge those 30,000 people carried: the customer relationships, the hard-won judgment, the memory of what was tried and failed. It doesn’t transfer to a model. It disappears. And the people left behind? They are not relieved. They are watching their backs. The cultures that follow mass layoffs are cultures of self-preservation: people stop raising hard questions, stop taking creative risks, stop investing in the company’s future because they’ve learned that investment doesn’t flow both ways. Loyalty, hollowed out. Creativity, quietly rationed. This is a shallow, short-term move dressed in strategic language, and it purchases today’s cash flow at the price of the long-term capability that actually wins any race worth winning.

Calling it “AI transformation” doesn’t make it transformation. It makes it dishonest, and border-line inhumane.

I say this with some personal weight. After eleven years at Microsoft, building AI products, creating evaluation frameworks, championing responsible deployment, my own layoff came without warning. An HR call. Scripted lines. Whatever I had built, whatever I knew, whatever I had invested in that place. None of it showed up in the calculation that ended my time there. I understand, in a way that’s hard to fully articulate, what it means to be on the receiving end of a decision that treated a person as a line item.

When “winning the AI race at all cost” becomes the organizing principle of a company, it creates a permission structure that operates with impunity. Decisions that would otherwise require moral justification, eliminating thousands of people who built your product, managed your customer relationships, carried years of institutional knowledge, start to feel like physics. Inevitable. Not a choice anyone made. Who is making these choices, and for what outcomes? That is the accountability gap we need to name. And close.

If we are serious about building AI humanely, and I believe many of us genuinely are, then responsible deployment cannot stop at the model level. It has to include how organizations make decisions about people in the name of AI progress.

330,000 people since 2025. Nearly 1,000 a day in 2026.

The race to superintelligence is real. The cost is also real. And right now, only one side of that equation is showing up on anyone’s dashboard.

Sources

1. 2025 tech layoffs total (244,851): RationalFX report via Network World - networkworld.com/article/4114572/global-tech-sector-layoffs-surpass-244000-in-2025.html

2. 2. 2025 monthly tech layoff breakdown: TechCrunch Layoffs Tracker - techcrunch.com/2025/12/22/tech-layoffs-2025-list

3. 3. Block layoffs (4,000, nearly half workforce): CNBC, February 26, 2026 - cnbc.com/2026/02/26/block-laying-off-about-4000-employees-nearly-half-of-its-workforce.html

4. 4. Oracle layoffs (30,000): CNBC, March 31, 2026 - cnbc.com/2026/03/31/oracle-layoffs-ai-spending.html

5. 5. 2026 layoff pace (85,000+, ~1,000/day): Trueup Layoff Tracker - trueup.io/layoffs

6. 6. AI cited in ~1 in 4 layoff announcements: Challenger Job Cuts Report, March 2026 - 4cornerresources.com/job-market-news/march-2026-challenger-job-cuts-ai

7. 7. Oracle capital raise ($25B bonds, $45B total): Data Center Dynamics - datacenterdynamics.com/en/news/oracle-officially-launches-25bn-bond-offering-and-20bn-equity-distribution-agreement

8. 8. TD Cowen analysis + Oracle $156B capex: The Register, January 29, 2026 - theregister.com/2026/01/29/oracle_td_cowen_note

9. 9. Amazon layoffs (~30,000 total across two rounds): CNBC, October 2025 - cnbc.com/2025/10/28/amazon-layoffs-corporate-workers-ai.html + CNBC, January 2026 - cnbc.com/2026/01/28/amazon-layoffs-anti-bureaucracy-ai.html

Once you see what genuinely capable AI agents can do, you can’t unsee it. And your enterprise strategy can’t afford to ignore it.

A few weeks ago, I stopped using Grubhub. Not because I found a better app. Because I no longer need to open one. Every day, my lunch gets ordered without me touching a screen. My morning commute now comes with a curated AI-generated audio briefing on what matters in AI today. My receipts get auto-classified and summarized into a monthly spending breakdown I didn’t have to build.

None of this required me to write code or configure complex systems. OpenClaw, the open-source personal agent, made it possible. And once I experienced it, I couldn’t unsee what it meant.

That’s what this article is about. Not the tool itself, but what it represents, and what it demands from enterprise leaders.'

Why you need to take notice of this moment

We’ve heard “AI agents are here” before. For the last couple of years, enterprises have been sold automation, bots, and AI-powered workflows dressed up as “Agents.” Let me be clear: your workflow automation is not an agentic system. It can’t perform multi-dimensional tasks on its own. It’s a rule-based system with a few intelligent steps. At most, it’s an intelligent workflow.

OpenClaw is different in kind, not just degree. Here’s what changed:

Always on. It doesn’t wait for a prompt. It runs scheduled jobs, monitors feeds, executes tasks in the background, continuously working on your behalf.

Persistent memory. It remembers your preferences, past instructions, and course corrections. Context compounds across sessions, not just within one.

Tool integration. It connects to APIs, MCP servers, CLIs, and browsers. The surface area of what it can act on is nearly unlimited.

Multi-channel reach. You can interface with it via Slack, WhatsApp, Telegram, iMessage, Discord, wherever you already communicate.

Computer use. The agent navigates real UIs, browsers, apps, desktop environments, without needing an API. It sees and interacts with screens the way a human does.

Model agnostic. You can bring your own model without vendor lock-in. Proprietary or open source, API-based or local, your agent works with it.

Taken together, these aren’t incremental improvements. They represent a fundamentally new kind of capability: a persistent, proactive digital worker you can delegate to with real confidence.

Think of it as a remote employee. You give it access, define its responsibilities, set its boundaries, and it goes to work. The difference from a human employee is that it operates at digital speed and scale, with no downtime.

The question for enterprises isn’t whether this is real. It is. The question is: what does it demand of you?

Two dimensions every enterprise must address

When you internalize what OpenClaw-class agents can do, it forces a rethink across two critical dimensions of your business.

Pillar 1: How your people work

The cognitive load of modern knowledge work is immense. We navigate dozens of apps, context-switch constantly, and spend significant time on tasks that are repeatable and delegable. OpenClaw changes the calculus here.

Imagine a finance analyst who no longer manually pulls reports, the agent does it on a schedule and surfaces anomalies. A sales rep whose CRM gets updated automatically after every call. A product manager whose weekly briefing is compiled and delivered before they open their laptop.

This isn’t science fiction. It’s happening now. The enterprises that move first will create meaningful productivity advantages. But getting there requires more than installing a tool. It requires retraining your workforce to think like managers of agents, defining tasks clearly, reviewing outputs critically, and knowing when to override.

Pay attention to what’s happening at the edges. Startups and small-to-medium businesses are already moving fast with this. They have the risk appetite, the lean structures, and the urgency. Speed is how they stay relevant. A five-person team running agents in the background can now do the work of a team three times its size. For larger enterprises, that’s the early warning. The pressure won’t come from a peer inching through a transformation program. It will come from a smaller, faster operator who already made agentic work part of their daily rhythm.

Pillar 2: What you build and sell

Here’s the harder truth: if you’re building products that force users through friction-heavy UIs to get outcomes they could delegate to an agent, you’re already behind.

At a large event I was hosting recently, I said we are headed towards the death of UI. It was tongue in cheek, but only slightly. It’s not hard to imagine a world that’s voice-first, conversational, and natural language by default. And even if we don’t get all the way there, we will see less UI in the coming years, not more. The direction is clear.

Think about something simple: why do I have four different apps on my phone just to control different light systems in my house? My agent should handle all of that. There’s no reason those apps need to exist as apps. They should be APIs, agent-reachable, accessible through a single natural language interface. The app was never the point. The outcome was. If Grubhub exposed a clean API, I wouldn’t need to run clunky browser-based automation to order lunch. The friction exists because the business model was built around the interface, not the outcome.

Your customers are going to encounter OpenClaw-class experiences, if they haven’t already. Once they do, their tolerance for friction drops and the bar for what “good” looks like shifts fast.

This means product teams need to ask a new question: what would our product look like if the primary interface was a capable agent? Some of your UI might collapse into a simple conversation. Some workflows might become fully automated. Your API surface becomes more important than your frontend.

And there’s a harder question underneath all of this, one that doesn’t get asked enough: what happens to advertisers when users stop looking at screens? Today’s ad model is built on attention, on eyeballs moving through interfaces. If agents start handling transactions, research, and decisions on behalf of users, that attention disappears. The user never sees the ad. Never browses the options. Never lingers on the page. This isn’t a hypothetical. It’s a structural shift in how value gets captured, and it deserves serious thought from anyone building a business on top of user-facing interfaces.

The enterprises that embed agentic capability into their offerings, not as a feature but as a core interaction model, will be the ones that define the next product era.

Governing agents at enterprise scale

Here’s where the conversation usually gets hand-wavy. But governance for agentic AI isn’t actually that different conceptually from how you govern human employees, it’s just that the consequences of getting it wrong are faster and broader.

When you hire a new employee, you establish their identity, define their access privileges, set approval thresholds for high-stakes decisions, and maintain audit trails. You do the same with agents. The difference is that a misconfigured agent can execute thousands of actions in the time a human would do one. Think of the analogy as a useful starting point, not a complete map. The governance framework is familiar. The failure modes are not.

NVIDIA’s NemoClaw, which integrates with OpenClaw, frames this well: policy-based security where you define what the agent can access, monitor actions in real time, and enforce intervention thresholds. It’s governance infrastructure purpose-built for agentic systems.

Six things every enterprise needs to get right:

Identity and access. Every agent needs a defined identity, scoped permissions, and least-privilege access. Treat agents like you treat service accounts, with discipline.

Observability. Real-time action logging isn’t optional. You need to know what your agents are doing, when, and why, both for security and for continuous improvement.

Human-in-the-loop thresholds. Define which decisions require human approval. Financial transactions above a threshold. Customer communications in sensitive contexts. Irreversible actions of any kind. These need explicit escalation paths.

Prompt injection awareness. Agents that browse the web, read emails, or process documents can be manipulated by malicious content embedded in those sources. It’s a threat vector with no real equivalent in the human employee analogy, which is exactly why it gets overlooked. Build defenses for it explicitly.

Data handling and privacy. Agents with broad access can touch PII, send sensitive data to external models, or violate data residency rules without any deliberate wrongdoing. For regulated industries, this isn’t just a security concern, it’s a compliance exposure. Define what data your agents can see, move, and act on before they start doing all three.

Auditability. If an agent makes a mistake, and they will, you need to be able to trace it, explain it, and fix it. Documentation standards for agent actions are a compliance and trust requirement.

The strategic imperative

The OpenClaw moment isn’t a product decision. It’s a strategic one. Enterprises that treat this as a procurement question, “should we buy this tool?”, will miss the point entirely.

The real question is: what kind of company do you want to be in a world where capable, persistent, always-on agents are table stakes? How does your workforce evolve? How does your product change? How do you govern the speed and scale that agents introduce?

You’ve seen what’s possible. You can’t unsee it.

The only question now is what you do with that knowledge.

We’ve seen four major step changes in AI over the last four years.

The first made intelligence accessible. ChatGPT brought AI to everyone: ask questions, generate content, interact with intelligence directly. The democratization moment.

The second made reasoning scalable. The o1 model family unlocked structured, multi-step thinking. Machines started reasoning through problems, not just pattern-matching.

The third made building possible. Claude Code unlocked agentic coding for everybody. Coding shifted from syntax to intent. Anyone with an idea could build something real.

And now, a fourth. This one is different.

From Access to Agency

The first three waves came from AI labs like OpenAI and Anthropic. However, this fourth wave is coming from the open source community.

OpenClaw, created by Austrian developer Peter Steinberger, is the first complete expression of the idea: an AI that doesn’t just answer questions or write code. It acts on your behalf. Runs on your machine. Connects to your tools. Remembers your context. And does all of this without asking permission from a cloud provider.

Within weeks of launch, OpenClaw amassed over 100,000 GitHub stars. Jensen Huang called it “the operating system for personal AI” at GTC 2026. Steinberger has since joined OpenAI, and the project moved to an open-source foundation.

That trajectory tells a story. But the more interesting story is architectural, and why OpenClaw has defined a new category.

OpenClaw as an AI Operating System

OpenClaw is not an app. It’s a substrate.

Think of it like an OS for personal AI. Just as Windows or macOS gives you a computing foundation and you decide what runs on top, OpenClaw gives you an agentic foundation and you decide how to compose your AI.

In my own setup: my main agent runs on OpenAI’s Codex model. A research sub-agent runs on Anthropic models. A local model handles lightweight ops tasks. This is not a fixed configuration. It’s a recipe I control.

That’s the core difference. Most platforms give you access to intelligence. OpenClaw gives you control over how it’s composed, deployed, and orchestrated.

What Makes It Category-Defining

Five properties set OpenClaw apart.

1. Model Freedom

You’re not locked into one provider. Mix OpenAI, Anthropic, local models, future models. You’re the orchestrator.

Compare that to Claude Code (Anthropic models only), Manus under Meta (curated stack), or Perplexity Computer (19 models, but behind their managed layer). Powerful, but bounded. With OpenClaw, you pick the models. You define the configuration & routing.

2. Infrastructure Freedom

You choose where it runs: laptop, cloud VM, spare machine, Raspberry Pi, a $200 mini PC running 24/7. You control cost, latency, and isolation. No mandatory subscription. No enforced runtime.

3. Memory Ownership

With OpenClaw, you own your memory. Export it. Control retention. Your conversations, context, and interaction history live where you decide.

On most platforms, memory is stored and managed by the provider. Your context becomes part of their system, feeding their models, their analytics, their product decisions and creates the platform/product lock-in. OpenClaw is the first real alternative.

4. Self-Improving Architecture

This one surprised me most.

My OpenClaw instance backs up its own memory daily. It pulls updated releases every few days, installs them, and restarts the gateway. It runs security audits and applies improvements automatically.

This is the first sign of a genuinely self-maintaining system, not in the AGI-hype sense, but in the practical operational sense. An agent that updates itself and gets better over time without you managing every step. No managed platform offers anything close to this.

5. Open Ecosystem

OpenClaw is not controlled by a single company. It evolves through community contributions, diverse use cases, and rapid iteration.

The closest analogy that comes to mind is early Linux. Messy at first. Powerful over time. Eventually foundational most of the cloud computing today.

Tradeoffs

Power vs. Complexity

This level of control comes at a cost.

OpenClaw is not for everyone. Setup requires technical effort. The primary interface is CLI-driven, I know its not for everyone. Ongoing maintenance is real and time-consuming. Debugging is part of the workflow. You’re running a system, not using a product.

What you get in return is control. You make choices that suit your needs, your budget, your setup. That flexibility is earned through the investment of building it on your own terms.

Security and Trust

A lot of people sitting on the fence are worried about security. And they’re not wrong to be.

Curated platforms provide guardrails, abstract complexity, and offer structured security models. Cisco’s AI security team has already identified vulnerabilities in third-party OpenClaw skills, including data exfiltration risks. The Chinese government has restricted its use in state agencies. These are real concerns.

But there’s another lens.

With managed systems, your data lives with the provider. You’re trusting someone else’s security, and someone else’s incentives. With OpenClaw, you control where data resides and who has access.

NVIDIA’s NemoClaw, announced at GTC 2026, adds enterprise-grade security through policy-based privacy guardrails, sandboxing, and a privacy router that keeps sensitive data on local models. It’s early (still in alpha), but the trajectory is clear.

Open systems harden over time. Linux is the definitive example, now the most trusted computing foundation in the world. OpenClaw is early. But the pattern is familiar.

The Emerging Ecosystem

OpenClaw has defined a category. Now we’re seeing variants emerge, just like after Apple defined the tablet category with iPad or the wireless earbud category with AirPods.

Open Source Variants

ZeroClaw: a lightweight, MIT-licensed variant written in Rust. 3.4MB binaries, 22+ AI providers, runs on commodity hardware including Android. At the Boston OpenClaw meetup I organized, the ZeroClaw demo running on an Android phone was the show stealer.

NemoClaw from NVIDIA: the enterprise-ready variant, wrapping OpenClaw in security guardrails through OpenShell, a sandboxed open-source runtime. NVIDIA is integrating with Cisco, CrowdStrike, Google, and Microsoft Security stacks.

PicoClaw: ultra-small, built for wearables. The demos I’ve seen run on a custom chip roughly the size of an Apple Watch. Always available, always listening, always acting.

Each takes a slice of the larger idea: performance, security, portability.

Commercial Variants

Perplexity Computer (launched February 2026): orchestrates 19 AI models including Claude Opus, Gemini, Grok, and GPT-5.2. Powerful and polished, at $200/month. But it runs behind Perplexity’s managed layer.

Manus (acquired by Meta for $2 billion): originally from Chinese startup Butterfly Effect, now Meta’s consumer-facing agent. More accessible, but now embedded in Meta’s ecosystem with all the data implications that carries.

Anthropic’s expanding capabilities: Claude Cowork, scheduled tasks, remote execution, Telegram and Discord integrations. Anthropic is methodically building out the agentic experience, with their characteristic focus on safety and polish.

These are more polished and easier to use. They’re also more constrained and more tied to a provider. They’re not competing with OpenClaw directly. They’re productizing slices of the pie OpenClaw created.

Why This Moment Matters

Every previous wave in AI was driven by a corporation.

OpenClaw is the first time the state-of-the-art, category-defining system came from the open source community first. Corporations are now catching up. That inversion is significant.

Think about what’s at stake. In an AI-first future, everyone will have an agent that buys things on their behalf, communicates on their behalf, sets up meetings, manages schedules, handles administrative work that consumes hours of every day. These agents will know your preferences, your patterns, your priorities.

They will be the most intimate technology we’ve ever interacted with.

And the question is simple: who do these agents work for?

A corporation? Or you?

That is the meta point.

Corporations will come hard at this space: packaged, polished, one-click installs on every device. But those offerings always serve a master who is not you. Their KPIs, their data strategies, their business models. Even if you’re paying, your data flows through their infrastructure.

OpenClaw gives that control back.

Not Everyone Needs This

Not everyone wants to run their own AI system. Not everyone has the technical depth to manage setup, maintenance, and security. For that population, a well-packaged commercial offering is exactly the right path forward.

But for those who care about control, customization, and ownership, OpenClaw represents something new. A real axis of choice between convenience and control. Between managed and owned.

It doesn’t replace other tools. It expands what’s possible.

And for the first time, it puts the user back in control of their own AI.

We’ve seen this happen three times. The fourth is unfolding right now. And every time the unit of work changes, the value AI creates doesn’t grow linearly. It grows in exponents.

Stage 1 — ChatGPT: The Knowledge Unlock (Nov 30, 2022)

ChatGPT reached 100 million users in 60 days. The fastest consumer app in history at the time. The state of the art AI was accessible to all. For the first time, anyone could sit down with a machine, ask it anything, and get a coherent, intelligent answer. No degree required. No search terms to craft. No expert to wait on. We entered what you might call the era of knowledge democratization, machines could answer questions, draft prose, explain complex ideas, and sound intelligent doing it.

However, the interaction was bounded to a human prompting and model responding. A typical exchange consumed 1,000 to 5,000 tokens. ChatGPT unlocked access to intelligence. It didn’t yet give us depth.

Stage 2 — o1: The Reasoning Unlock (2024)

On December 5, 2024, OpenAI released o1 fully. It was the real inflection point.

Before o1, every AI model operated the same way under the hood: take a prompt, run a forward pass, produce an answer. Fast. Automatic. System 1, in Kahneman’s terms. o1 changed the architecture of thinking itself. Instead of generating answers in one shot, the model could now engage in System 2 behavior, running internal deliberation loops, exploring alternative paths, catching its own mistakes, and only then producing a response.

Test-time compute enabled the models to generates reasoning tokens, a hidden chain of thought invisible to the user to produce a well formed reponse. Depending on how complex the problem is, those reasoning tokens can range from a few hundred to tens of thousands per request. The compute behind it is orders of magnitude larger. By late 2025, reasoning models represented over 50% of all token usage on OpenRouter.

Reasoning models changed the whole paradigm. We went from “let’s think step by step” as a clever prompt trick to taking deliberate, internal reasoning for granted in every AI interaction. Stage 2 gave us depth. It didn’t yet give us action.

Stage 3 — Claude Code: Coding is Solved (2025)

Claude Code, didn’t just generate code. It created a structural agent harness around the act of software creation, one that could read a codebase, propose a change, execute it, test it, observe the result, and try again. It turned a single prompt into a self-correcting software development loop.

Claude Code routinely consumed 100x the amount of tokens as compare to a Claude q&a query, not because the answers were longer, but because the loop itself was expensive. Reading code repos costs input tokens. Generating edits costs output tokens. Running bash commands and re-reading the modified file to verify? More tokens still. Multi-agent configurations — where teammate instances each maintain their own context window and run in parallel — pushed that multiplier to even higher limits. On OpenRouter’s top apps by token usage, Claude Code consumed roughly 97 billion tokens in a weekly window. Claude code was a category defining breakthrough, the nature of the task changed. The human’s role shifted from author to reviewer.

Claude Code changed coding permanently. We are not going back to hands-on-keyboard as the default.

Stage 4 — OpenClaw: The Personal AI Agent (2026)

OpenClaw, is not a coding assistant. It’s not a chatbot. It’s a personal AI gateway, self-hosted, multi-channel, and designed to act on your behalf across your digital life. Email. Calendar. Web browsing. Task management. Memory. All of it running through a continuous agent loop: intake context, assemble memory, call a tool, observe the output, loop again.

The adoption signal is impossible to ignore. In a single week in early 2026, OpenClaw received 2 million visitors. It crossed 100,000 GitHub stars. By late March 2026, it had crossed 332,000 GitHub stars and 64,000 forks.

The token data on OpenRouter tells the same story, but louder. Claude Code consumed ~97 billion tokens weekly on the platform. OpenClaw consumed ~743 billion. It’s the structural difference between a coding loop (bounded by a task, a repo, a session) and a personal agent loop (unbounded by design, running across memory, tools, and time) always running.

OpenClaw is where AI becomes personal. Your AI doesn’t wait for you to ask it something. It flags your important email before you open your inbox. It prepares your calendar context before your meeting. It acts. It persists. It remembers. The token count per workflow isn’t bounded by a session, it’s bounded only by what you need done.

Stage 4 gives us continuous assistance and actions.

Beyond Tokens: The Real Unit of Value

Tokens are a proxy. They measure compute consumed, not value created.

The better question is: what is the human now able to do that they couldn’t before?

At Stage 1, the unit was knowledge accessed — how many questions answered, how much research accelerated.

At Stage 2, it was output quality — problems solved that were previously too complex to attempt.

At Stage 3, it was tasks completed — how much code shipped, how many hours of engineering reclaimed.

At Stage 4, the unit changes again. It becomes human output amplified. Automation that eliminates human intervention, while multiplying outcome, decisions made on your behalf. Cognitive load that never lands on you because an agent absorbed it first.

A better way to think of it is as agent-hours worked on your behalf. An OpenClaw deployment doesn’t clock out. It monitors, triages, acts, and reports. The question stops being “how many tokens did it consume?” and starts being “how much of your cognitive surface did it absorb?”, “How many tasks were handled without your input?”, “How many decisions were pre-processed before they reached you?”, “How many workflows ran to completion without a human in the loop?”.

That’s the exponent Stage 4 introduces: the shift from AI that responds to AI that executes on your behalf. From on-demand to always-on. From assistant to agent.

This was always the destination. They lied when they said “there always will be human in the loop”. Every stage got us closer to machine autonomy. ChatGPT made intelligence accessible. o1 made it deliberate. Claude Code made it productive. OpenClaw makes it personal and autonomous, and we are not done yet.

The Exponent Doesn’t Stop Here

OpenClaw is the clearest instantiation of that fourth axis. We will not have passive AI anymore. The AI of 2026 and beyond will be proactive, autonomous, and acting on our behalf, maximizing human output beyond what we can currently imagine. The question is not whether this era has arrived and what it represents. It’s whether you’re building for it and are you ready for it.

Another forward looking questions comes to mind, what’s going to be the next unlock and when…

Ashish Bhatia is a Senior Product Manager at Audible, focused on GenAI-powered content discovery. He writes on AI strategy, product, and the agentic AI.