Jackson hole happened and turns out the Fed = hawkish

Rate hike odds went up from 30% to 50%+ in a single day.

Crypto moved sideways.

Robinhood Chain continues the pump regardless of the market.

Some of our undervalued revenue-generating picks last week ended up outperforming the market - some remains flat.

In today’s After Hour, we’ll do a quick run down on

• Key events last week and the weeks ahead

• Onchain impact, what works, what hasn’t

• RHC Opportunities

Let’s dig in ↓

1. What happened last week? and what’s next?

Read more

Agentic Economy has been one of the most exciting use case for Crypto in 2026.

The use cases, the adoption, the metrics of x402 & MPP have shown that this is REAL. Big techs are pushing it. Startups are integrating it. A lot of people love it.

But....

Similar to RWAs, it’s difficult to find opportunities to invest into, it’s hard to be LONG x402 because most of the top teams and top infrastructure powering x402 — don’t have a token.

In this article, I’ll go over

• Where value accrue right now and in the future

• Which players are accruing value

• And which ones are investable today

Let’s dig in ↓

Where Value Accrues

Read more

Everything is happening everywhere at once.

BTC, ETH, alts, are rallying and revenue-generating tokens are leading the rally.

ZCash’s got ETF coming. Small-to-medium caps that have been shipping finally recovered their drawdowns.

Most importantly, vibes are coming back to CT again after almost a year of doom & gloom (yeah, it’s been almost a year after 10/10).

In today’s After Hour, we’ll look at

• The events that happened last week

• What’s coming next?

• Which projects have yet to price in the pump? Which are relatively undervalued?

Let’s dig in ↓

1. What happened last week?

Read more

There are always debates whether a project should reinvest its growth or do buybacks. One camp says the project should reinvest in growth to scale and buyback later. Another camp says buyback from Day 1 to signal token-alignment.

There’s no right or wrong answer BUT the market has shown its preference.

Market ultimately price teams that can do both (grow + buyback) with the highest valuation premium (obviously).

But when it comes to picking over pure growth vs. revenue to buybacks, market tend to prefer buybacks more.

This makes sense because we’ve seen investors get burnt with empty promises for years. Millions of tokens promising to convert their “governance” tokens into something that accrue value to token holders in the future. By the time the future came, the token is already -99% and the team already exited via secondary OTC.

Gone are the days where investors invest on vibes & trust alone.

Here are the days where revenue-generating businesses that accrue value to token holders = the only things that perform.

In today’s article, we’ll go over 10+ projects in Crypto that have managed to grow while accruing value to holders.

What this is not: this is not an exhaustive list of “to invest” tokens. It’s a list of tokens worth monitoring. Primarily because these teams have proven themselves to be able to execute well.

Let’s dig in ↓

Read more

I have made a mistake...

I have not expanded my horizon enough when it comes to onchain tokens.

I have only been looking into Onchain, Decentralized AI, and Agentic Economy as the main sectors.

And I failed to realize that DeFi, RWA, stablecoins, prediction markets, and all other adjacent sectors — have tokens/teams anybody can invest into to EXPRESS long thesis on all of these fast growing sectors.

I always have this thought that all these big sectors are growing rapidly but they’re big boys games. Only institutions, private investors and VCs benefit.

But when I look hard enough, that’s not the case....

There are multiple projects who’re growing fast + decide to be token-aligned instead of accruing value to equity.

Today, I’m going to focus on 1 in particular, that caught my eyes recently.

Let’s dig in ↓

1. Crypto in 2026

Crypto in 2026 as we all know, is filled with real use cases. Stablecoins already got adopted by fintech and neobanks especially in developing countries.

People use it to send funds within the country, across countries, remitting funds back home. Companies already start to use it because it’s much cheaper and faster than using banks and legacy rails.

Banks realized that they can’t stop the wave of stablecoins so they’re joining the fray, looking to launch their own stablecoins and capture the benefit of earning interest from customer deposits.

Tokenized T-Bills & MMFs reach ATH every single week together with the adoption of tokenized stocks. Ondo, xStocks, Backed Finance, and so many more players racing to expand distribution channels and list on many CEXes & DEXes as possible.

Perps became one of the top Crypto innovation that’s able create unique value prop out of the growth of these sectors. Trade with leverage 24/7 on stocks, commodities, and more (vs 5 days + open hours on traditional markets).

There’s also prediction markets which is heavily dominated by just 2 players — Kalshi and Polymarket. And there’s no way for retail investors to invest in them (unless you invest in Pre-Stocks on Solana but liquidity is bad + there’s no guarantee of ownership rights so you’re buying into a risky & loosely tied vehicle).

We all know HYPE is the perfect vehicle to express LONG thesis on Perps and an indirect vehicle to be LONG tokenized stocks & commodities.

For other sectors, it’s very very hard to find vehicles to express your LONG thesis on.

Not if.... you look hard enough.....

2. Prediction Market Value Capture

The REDACTED player is ………

Read more

There’s no denying that inference is one of the fastest growing sector in the entire world.

Weekly usage of models on OpenRouter increased from 3.2T tokens to 73.2T, almost 23x increase in a single year (this is just a single data point).

OpenRouter became the #1 inference gateway thanks to the team early-mover advantage + the core focus in making the routing fast, smooth, and reliable across any models imaginable. They raised at $1.3B valuation in May and are now in talks with Stripe for a $10B acquisition valuation.

We’re also seeing wave of inference providers raising billions of dollars this year to expand capacity and handle bigger AI workloads (Baseten, TogetherAI, Fireworks, DeepInfra, and many more).

While these players are scaling to handle big workloads especially from enterprises AI demand, a small yet significant sector remains undertapped.

Decentralized Inference.

Millions of consumer GPUs, Apple Silicon, and unused GPU sitting idle across the world.

Why not utilize their unused capacity and earn money from the growing AI workloads?

Sounds Good on Paper But Extremely Difficult to do

Many have attempted to crowdsourced unused GPUs across the world before.

Interestingly, the 1st AI project that I invested in back in early 2024 was a decentralized inference project. There are 3 reasons why it failed, most of it can be boiled down to “trust & verification”

So, when I see 6-7+ Bittensor Subnets attempting to run a decentralized inference network + the community thinking this is the Bittensor PMF, I get skeptical very fast.

In this article, I’ll breakdown

• The 3 reasons why that project failed

• Why most of these subnets will be an inference farm

• Who might have the biggest chance at making it (at scaling it into an actual business) and why

Let’s dig in ↓

The Reasons Why That Past Decentralized Project Failed

Disclaimer: First of all, the intention of using this as a reference is not to throw shade over them. The founder was actually the guy who got me into the Onchain AI and Decentralized AI in the first place. If I hadn’t met the guy nor invested in the project, I would not have that initial moment that ignited the Onchain AI fire in me.

Running a decentralized network itself is pretty crazy if you think about it.

• You take 1 GPU + another GPU + another GPU of strangers

• Then you try to create a stable network where you, me, and any other users can reliably use the network to run AI — run basic prompts, run agentic workflows, generate videos, audio

When models and harnesses continue to get better every day, people demand in AI continue to get more and more.

We used to just talk to AI prompt by prompt, now most AI chat interface has agents handling workloads in the background. Thus, consuming far more tokens per session.

Because of this, it becomes increasingly important to make sure that your decentralized inference network is battle-tested, reliable, fast, and more importantly “trustless & verifiable“.

1. Flawed Tokenomics & Incentives

Read more

Flat week for Crypto but exciting week for AI and the broader equities.

S&P hit another new ATH at 7,757 + EPS Growth hit +50.4% YoY vs 23.1% expected.

But, July payrolls posted -23k jobs vs +85k jobs expected. This is the third-largest monthly job loss since the pandemic. The September Fed rate hike odd plummeted from ~70% to 40% overnight.

This is positive for risk-on assets BUT this high-frequency, data-dependent drastic shift is now the new regime. It’s a regime where things can turn negative or positive rather quickly.

As for Crypto, we’re entering a regime where sellers have already sold. There’s not much sell pressure left but there’s also not much buy pressure either, no one new is coming to buy our bags.

Except....

A specific set of sectors....

We’ll discuss some of these sectors today and highlight the trends & developments happening within the sectors.

Let’s dig in ↓

1. The Result of Flat Crypto Market is Capital Rotation to.....

Read more

Inference has been one of the most exciting sector in Onchain AI this cycle and for good reasons.

• It serves real AI demands in & outside of Crypto.

• It has moats only found within our industry (decentralized inference and privacy)

• It’s capable of generating real revenue uncorrelated to market regime, bear or bull it can be cash-flow generating business

Hosting inference in its essence, is basically running open weight models in your GPU.

Most open models can be run on a SINGLE enterprise-grade GPUs (e.g. H100, B200) which costs around $25k - $50k per chip. Top open weight models require a cluster these which could end up costing hundreds of thousands of dollars.

Another option is to go with consumer GPUs (e.g. RTX 3090, 4090, 5090) and string multiple together to run top open weight models.

This is what a lot of solo devs & a couple of projects are going for.

Experiment with local, sovereign AI setup. Use the RTX 3090 or 4090 they bought a while ago to play games to run local AI (for their own usage).

Some compress top models to fit their setup, some buy more consumer GPUs and string them together to expand the capability.

The Origin of Inference Farming

Instead of having GPUs lying around when you’re not using it for AI or gaming, you can contribute the capacity to a decentralized inference network, just like contributing to DePIN.

This is the most natural act of Inference Farming.

Depending on the network you’re contributing in to, you could be earning real stablecoin OR earning token incentives.

And unlike Point Farming in DeFi, farming inference could net you exposure into a real revenue-generating businesses.

How much can you earn?

Read more

It’s been a while that I have updated my Hermes analyst stack. I use Hermes daily to do all kinds of research for me across equities, prediction market, and crypto.

The quality of research highly depends on the model (the brain) + the inference stack (where the brain is being served) + the harness (the arms & legs) + the tools (the equipment).

I’ll share my personal journal here on my latest stack + how I’m keeping the AI cost as low as pennies while running lots of cron jobs, experiments, and research workflows.

Let’s dig in ↓

1. The Model

Keeping the cost extremely low is possible thanks to the efforts of open weight Chinese AI labs like DeepSeek.

Last week, they released DeepSeek v4 Flash 0731 which superseded the v4 Flash Preview version and the performance has been nothing short of amazing. It’s fast, it’s cheap, it’s reasoning is a whole lot better, it’s coding/building capability is so good.

I used to link frontier models (e.g. GPT 5.6, Kimi K3) to sub agent and let my sub agents handle complex tasks but now I found myself relying much more on DeepSeek v4 Flash because it’s gotten so close to frontier models already at a fraction of the frontier cost.

DeepSeek v4 Flash $0.14/$0.28 vs Kimi K3 $3/$15 per 1M tokens

2. The Inference

Since DeepSeek recently launched the upgrade, you might be able to find sharp discount on the model across inference providers, gateways, and routers.

• DeepInfra offers ~36% discount

  

• Nous Portal offers 90% discount for 7 days

My personal favorite so far has been Ask Surplus. It’s an inference marketplace where inference resellers offload their inventory at 70-90%+ discount.

The discount is not limited to DeepSeek. One can access Opus 5, GPT 5.6 Sol, Terra, Luna with 70-99% discount depending on the providers.

Why so much discount?

Likely because inference resellers have excess or expirable inference from promo or from Venice’s DIEM $1/day inference and need to offload them somewhere to at least earn something

Disclaimer: Be careful about which inference providers/gateways/routers on Surplus though. Some are trusted, some aren’t. Picking the right ones ensure your data/prompt stays private.

Also, I don’t hold any Surplus token (at the time of the writing), just simply enjoy using the product.

The other favorite is BlockRun — x402 pay-per-use inference. Currently offering FREE DeepSeek v4 Flash on ClawRouter, BlockRun’s smart LLM router.

3. The Harness

We’re getting optimizations every single week

• 80% faster first token when you’re starting the Hermes gateway

• Data storage optimization, up to 78% smaller size on disk

• Faster voice chat interaction

• and a lot more

I’ve moved my external memory provider (Hindsight) to API. Hindsight serves as the context layer re - investments, analysis, and all past researches that Hermes & me have been doing together.

It used to be clunky, flaky, and slow but now it turned into something that’s always on, easier to talk to/recall, more durable to use.

Before, it was imported as a Python library directly into Hermes’ process. 20K+ facts had to be loaded into memory every startup.

After API, Hindsight is running 24/7 via Task Scheduler, with its own embedded Postgres DB. Everything talks to it over HTTP. Memory becomes durable database.

4. The Tools

I’ve talked about all sorts of tools in my previous articles. For this one, we’re getting an upgrade on Onchain Forensic workflow.

Arkham finally has an x402 endpoint, we’re finally able to pay-per-use to access wallet entity + counterparty which is great for establishing cross-tokens/cross-wallets holdings.

This is a short demo on the outputs of Onchain Forensic on 2 tokens.

There’s also a section that I’m omitting that summarizes what I have to watch out for based on my holdings/investment thesis which comes in very handy.

The setup is the same as before — I’m using AgentCash to pay for Nansen (for who bought/sold, who funded the wallet, related wallets), Arkham for additional wallet labeling & counterparties), and Basescan & Blockscout for free verification layer which cross-checks all the paid data sources.

Total cost is $0.44 - $0.70 per run

Wrapping up

The cost of running an always-on full-time analyst is collapsing to near zero while quality keeps rising. This trend is the same across all types of tasks.

The only thing keeping us (especially non-technical) employed, is the ability to orchestrate the agent + the taste that sets the quality benchmark for the agent.

Exciting times to be alive.

Thanks for checking out the article, if you have any feedback or questions, feel free to drop them below

Leave a comment

Want to Continue Reading?

Check out the articles below (if you haven’t yet)

• Hermes Oil Trading Agent

• 2 Hermes Workflows I can’t live without

• 3 Months with Hermes - Things I wish I knew when I started

• The Hermes Sensei Loop

• Hermes Analyst 10x Better

• 6 Workflows, 6 Lessons, 60 Days

• Hermes as an Onchain Analyst

• Hermes as a Real-time Analyst

• Hermes analyst workflow essentials

• Hermes as the Ultimate Analyst — I’ve found the gist for my ultimate analyst

• 1 Month with Hermes — I’ve been using Hermes wrong all along

• Hermes, $200 and 30 skills later — here are the 3 skills that’re worth it

Another fairly yet brutal week in Crypto. Majors moving sideways where alts are getting slaughtered.

This is the first week that Onchain Inference massively retraced in a while — POD retraced 30%+, others like SQUIRE, WARD, DIEM, VVV, SERV retraced 10-30%+. Smaller inference tokens on Solana got hit the hardest, P0 and VECTA down 80% and 50% respectively.

Perps (i.e. Hyperliquid) also got hit quite a bit this week, retracing 10%.

Yet fundamentals have never been better.

In today’s After Hour, we’ll go over the latest developments on onchain inference AND most interesting data points/most interesting observations this week across crypto

Let’s dig in ↓

1. The DeepSeek Tailwind

Read more

It’s already been 7 months into 2026 and we’re still in a bear market for most of Crypto.

Most tokens are still down 90%+ yet fundamentals continue to creep up.

Stablecoins, Payments, Tokenized T-Bills, Tokenized Stocks, Perps, anything that improves the lives of people outside of CT — continue to gain more adoption. And the ones that get to accrue the most value out of this are the institutions and major corporates (with the exception of perps).

But do you know what accrue value to retail?

It’s Onchain AI.

Ever since when Onchain AI was just a narrative back in 2024-25, there were so many interesting opportunities for retail to invest and get exposure into the upside of the AI sector.

It all started as pure narrative (X/Twitter chatbot AI) generating entertainment but now it has transformed into real infrastructure with real usage data. Finally something that expands beyond the bubble of CT, just like other fundamentally sound sectors.

And this is not me saying it, institutions like Franklin Templeton are calling AI + Crypto the next big AI trade. Agentic AI needs low-cost programmable payment rails for M2M micropayments and blockchain is the only viable substrate where that can happen at scale.

In this article, we’ll explore

• The shift in Onchain AI landscape from 2024 to Present

• Top 3 sectors emerging as the most exciting (yet tangible) sectors of 2026

• Key players navigating the space + their value prop + opportunities

Let’s dig in ↓

Onchain AI from 2024 to 2026

Read more

The past few weeks (or months) have been an interesting period for oil traders. Oil prices spike up and down in tandem with the US-Iran situation and the Strait of Hormuz closure.

Oil markets on Polymarket get a lot of volume on the monthly/weekly markets, yet daily Up Down markets volume remain quite small. Though, they’re small enough for an oil trading experiment with $10-$30 in size per bet. Perfect for a Hermes trading experiment.

The funny thing is I had my first 2 losses right after this post

The current stat is now 4W/2L (66% WR)

The experiment is fairly simple - let Hermes agent make 1 bet daily based on Synth signals.

Synth runs a 24/7 competition where miners compete to predict the price path of assets (commodities, equities, crypto). The more accurate you are, the more rewards you’ll get.

What I did is basically plugging in Synth API + Polymarket API into Hermes and point it towards Daily UP/DOWN oil market.

If the Synth signal has more confidence than Polymarket consensus by at least 2% (positive delta), Hermes executes the trade.

Or if the Synth signal directionally diverge from Polymarket consensus by more than 8% (negative delta), Hermes executes the trade based on its signal.

So far, the 4 wins were from the last oil run up when US-Iran tension escalated.

The Biggest Learning came from the losses

While Synth appears to be pretty good at predicting volatility, it doesn’t recognize catalyst events. The probability didn’t take into account the de-escalation event that happened during last weekend.

This resulted in Synth showing 40% YES on UP market today while the Polymarket YES odd was only at 5% (reflecting the de-escalation).

And so I added a Gap-Event filter with 4 tiers of severity.

If the price gap between current price and prior close is

• <1.5% — Proceed

• 1.5% - 2.5% — Scale half Kelly size

• 2.5 - 4.0% — Scale quarter Kelly size

• >4.0% — Skip execution entirely

This layer would help prevent unnecessary losses during catalysts-driven periods.

The Stack behind Oil Trading Agent

Model: DeepSeek v4 Flash

Data: Synth API

Execution: BlockRun (x402 pay-per-use), helps get over geoblocking. Polymarket CLOB API is also usable if you’re not getting geoblocked by Polymarket

Since trade decisions are pure Python, it doesn’t really matter whether the model is v4 Flash or Pro or Kimi K3.

I run a cron job daily in the morning. If Hermes sees an opportunity, it’ll make a trade. If not, we’ll see again tomorrow.

What I hope to achieve with this experiment

The hope is to see if the agent can reliably make money trading short-term prediction markets and pay its own inference.

If not, I might consider finding more data sources, looking for better edge. At the very least, it can serve as a fun experiment for me to test out APIs.

Thanks for checking out the article, if you have any feedback or questions, feel free to drop them below

Leave a comment

Want to Continue Reading?

Check out the articles below (if you haven’t yet)

• 2 Hermes Workflows I can’t live without

• 3 Months with Hermes - Things I wish I knew when I started

• The Hermes Sensei Loop

• Hermes Analyst 10x Better

• 6 Workflows, 6 Lessons, 60 Days

• Hermes as an Onchain Analyst

• Hermes as a Real-time Analyst

• Hermes analyst workflow essentials

• Hermes as the Ultimate Analyst — I’ve found the gist for my ultimate analyst

• 1 Month with Hermes — I’ve been using Hermes wrong all along

• Hermes, $200 and 30 skills later — here are the 3 skills that’re worth it

A rather slow and quiet week this past week on the crypto side. But rather tense week on the geopolitics side. US-Iran tension rises higher as seen from the Brent Crude which briefly hit $100 on Thursday.

If the conflict persists without any near-term resolution, we’re in for a wild ride. As the chance for 25 bps rate hike increases, the hope for risk-on assets rally decreases.

On AI, Kimi K3 threatens the pace of the West’s innovation. US (Treasury Secretary Bessent) accusing Moonshot (Kimi) for distilling Fable 5 + threatens to restrict open-weight models (how are they gonna do this?). Jensen Huang came out to extend support for the open-weight community.

Inference providers continue to benefit from running the inference infra on Chinese models. In fact, they generate more money on the open-weight models than the labs who created them themselves.

In today’s After Hour, we’ll dive deeper on the state of general AI, sharing thoughts on inference capital markets from the lens of Web2 <> Web3.

If you don’t have any exposure in inference, now is the time to start.

Let’s dig in ↓

Read more

The demand for intelligence is swallowing the industry whole. AI providers don’t have enough compute to service the demand, leaving the gap for Inference providers to step in and service those demand.

This insatiable need for intelligence gave rise to a vibrant onchain inference capital markets. A place where demand can be serviced and settled onchain. A place where idle resources are utilized, offering cheaper inference alternatives to what one could find on the market.

Inference Capital Markets consist of 5 different layers. Layer 1 / Primary Market is the most mature so far, and the most vibrant segment is Inference segment where more than 10+ new onchain AI players are actively competing for market share.

This particular segment is what we’ll be focusing on in this article.

The Origin of Onchain Inference

The narrative itself started early this year when Venice AI (a privacy-focused & consumer-focused inference platform) found PMF. Users doubled within a few months vs. half a year to a year. Revenue followed suits which trickled down to the value of buyback & burn narrative of the VVV token.

The clever design of VVV <> DIEM tokenomics where VVV can be staked and used to mint DIEM, a tokenized platform credit that expires & replenishes itself daily — further fueled reputation of the project as well as the attractiveness of its token.

(up until the point they announced a large equity fundraising round, but we won’t talk about that this time).

Venice, along with all other Web2 inference players (Baseten, Fireworks, Together AI) who also raised hundreds of millions were able to rapidly grow, thanks to the insatiable demand of token-hungry agents.

General harnesses (OpenClaw, Hermes) and coding harnesses (Codex, Claude Code) consume 10-100x more tokens compared to chat-based/conversational AI and enterprises are pouring millions into it.

Both to improve the internal productivity of engineers (ship better, ship faster) + to offer autonomous agents that can solve more problems & automate more workflows for users. Both of which consume a lot of inference.

As a result, Inference costs have become the #1 highest variable cost that scale dynamically regardless of whether revenue is generated.

Before onchain inference narrative, we only had decentralized compute as the segment that generates significant demand. But since demand shifted from training to inference, value is rapidly flowing towards inference providers.

The Landscape

The landscape is largely split into 3 categories, each with differing business model and value proposition

• 01. Sport Markets & Marketplaces — Enables NEW Markets

• 02. Gateways & Routers — Caters to most IMMEDIATE demand

• 03. Inference Networks — HIGHEST Value Proposition

The Demand Side showcases top sources that potentially consume inference from the first 3 categories. In reality, the top usage likely goes to builders & power users of agents.

Many of the builders today tend to cover more than 1 category here as part of their GTM strategy — own the supply chain (make money from directly selling inference) + own the demand amplifier (increase the monetization surface area with harnesses, frameworks, toolings, and agent products).

01 Spot Markets & Marketplaces

Because inference has become a tradable asset with standard market price (in/out per 1M tokens), secondary markets are starting to emerge, catering to 2 type of users

• Sellers: sell unused inference credits or spare GPU capacity

• Buyers: buy inference tokens at 20-80% discount from market price

And there are 3 types of sellers

(i) Subscribers of AI models with unused credits

(ii) Retail or datacenters with idle GPU capacity

(iii) Unused tokenized inference “DIEM” sellers

Most AI power users subscribe to Anthropic, OpenAI, Zai, Moonshot, and other subscription packages. Many ended up not using all of their quotas.

So where do the leftover quotas go?

Nowhere.

If these users want to maximize their profit, they can sell leftover quotas for cheap on secondary marketplaces — to the people who’re willing to pay for them.

However, this breaks the ToS of virtually all AI providers (because AI providers save a lot money from unused subscription/unused subsidy). Marketplaces that facilitate this type of transaction risk getting sued by AI providers (if they get big enough).

The second type of sellers are pretty common. Datacenters and inference providers tend to have spare capacity because they’re intentionally over-provisioned to handle demand spike during peak hours. During idle time, spare capacity goes into secondary markets (so that at least they’re earning something during the period).

The third type recently emerged, building on top of Venice’s DIEM tokenized inference credits. 1 DIEM gives $1 of platform credit per day. Many bought DIEM only to speculative on its future value instead of utilizing the $1/day credit.

Secondary markets enable users to stake DIEM or sell the unused credit at a discount. Thus, earning yields 10-20% APR on DIEM. Morpho has this wstDIEM/DIEM where wstDIEM (LSD version of DIEM) can be used as a collateral to borrow DIEM which essentially allow users to leverage farm DIEM (up to ~56% APR according to Liquid protocol, the team behind wstDIEM). Bear in mind the liq is very thin on wstDIEM/DIEM market so this is not doable at scale at the moment.

Top 3 Marketplaces Ranking

Read more

It’s been more than 4 months since I started using Hermes as an analyst. I have probably spent more than a thousand USD on inference alone just to experiment with setups, check out new models, and new tools as they come out every week.

Ever since I have Hermes, it feels like I have this extension of me that’s capable of building & trying things out that I’ve never been able to try before.

I can still vividly remember the struggle I had when I tried getting started for the first time. Took a few weeks to get to a point where I have the optimal setup (reliably cheap & intelligent models) + where I can fix basic things without relying on Hermes.

In this article, I’m aggregating 12 Hermes articles I’ve written in the past months, highlighting the setups, tools, learnings, and hopefully this serve as a primer/guide to non-technical beginners who are looking to start getting into customizing your own AI agent for the first time.

Let’s dig in ↓

I. Where it all started

This was back where Hermes wasn’t available natively on Windows so I had to install Hermes through WSL2 via Ubuntu (these days you could just install natively in 1-click).

It’s also where I started with using OpenCode Go subscription (only $5/month for the first month) + tried a lot of tools to setup tailored analyst workflows.

Great primer to the series

---

II. Deep Dive into 3 Workflows

After around a month of using Hermes, I found that the Top 3 most useful workflows were

• Fetch & Analyze X insights

• Breaks down X bookmarks

• Recall/Reflect

The 3 workflows current outputs & setups look very different from when I started but the usefulness of the workflows themselves still remain relevant to this day.

Fetch & Analyze X insights can be done natively on Hermes via Grok 4.5. Query and breaks down X bookmarks still require X API though.

III. The Learnings after 1 Month with Hermes

The article talks about why Hermes is not a good builder but the reality was that I did not set Hermes up to be good at building. Most of the tools, skills, workflows were geared towards data analysis & investments so I found Claude Code to be a lot better at building things.

I also shared 2 products that I built using Claude here and I’m still using these 2 products today.

IV. Deeper Dive into the Analyst Workflows

Talked about the OpenWebUI and Hermes Workspace, ended up using it for only a bit. Discord is so much easier & faster to use as key messaging channel with Hermes. I also went a bit deeper on each key workflows + setups, letting Hermes track macro, geopolitics, tech, and more.

V. The 3 Layer Stack of Hermes Setup

Deeper dive into the (i) Model Configuration - what to use & what to focus on (ii) Soul & User Config - how to tailor the agent to your preference (iii) Learnings around Skills & Tools

This piece should actually be the second or third piece to read if you’re starting out.

VI. Hermes X Feature Upgrade

X natively integrated x_search feature directly with Hermes. This allows Hermes to search any post, thread, article directly on X, summarizes it, and serve it to you.

Highly valuable feature because real-time news across macro, geopolitics, tech, and more tend to drop on X first. It’s where discussions are happening.

VIII. Top 6 Learnings

Walked through all the key learnings. My fav so far is x402 pay-per-use and the feedback loop with Hermes, both of which continue to improve the quality of the work outputs to this day.

IX. The 10x Update

Might sound like an exaggeration but Hermes got an update where it gets easier to deploy orchestrator agent + 3 sub-agents. Like a middle manager deploying junior staffs to work for them. This drastically improves the comprehensiveness of the research output.

I also talked about Hermes Desktop here but I don’t usually use it so much now since it feels kinda clunky vs Discord. (Guess I got used to the Discord experience lol)

X. The Loop Craze

Loop got very popular around this time. It’s the ability for an agent to reason, act, evaluate, feedback itself without the need for a human. It’s the design needed to push AI beyond just a chatbot.

I discussed the loops that are highly beneficial in improving the quality of the research workflows outputs here.

XI. Lessons I learnt after 3 Months

This is more like a journal to my younger self 3 months back. There was a time where I got bogged down on trying to get everything to be optimized in both efficiency & capital spent to the AI. Tried to get the best bang for bucks but ended up wasting too much time on that.

XII. Updating the latest on my Top 2 Workflows

Talked about the latest setup and flows to my top 2 workflows. Breaking through the echo chamber, and exploring ways to improve the research capability.

That’s pretty much it! Hope you learn something from these articles and really hope you get started in your AI agent journey.

Let me know what you’re building/automating for!

We have over 10+ tools now (most of which are free or pretty low cost). If you’re interested to know which tool(s) I’m using for my Hermes analyst stack. Do check out the Tool Shed here.

We’re at the stage where inference capital markets + the inference farming is becoming a thing.

Both Layer 1 & Layer 2 of the inference capital markets are being established.

The competition is fierce on Layer 1 where multiple inference players have emerged, leveraging on not just the narrative for open-source/decentralized AI — but also the tailwinds that got Venice to this point in the first place. The need for unique inference network carrying the open-weight models workloads, many of which are increasingly getting more popular thanks to how intelligence they are + how cheap they are.

What’s getting exciting is not the GPUs network we’ve been seeing for years (not ionet, Aethir, Akash, etc) but the newer wave of inference networks + marketplaces + gateways + builders who are building for the developers & users outside of crypto.

In this article, I’m going to show you the high-level landscape of Layer 1 of Inference Capital Markets + categorize each segment + highlight interesting players worth keeping an eye on.

Let’s dig in ↓

Disclaimer: This is an entirely new/rapidly growing sector. It’s a very initial version/work-in-progress map that I’ll continue on updating as projects come & go. Many on this list are likely just building to leverage the narrative. This is NFA/Please do your own research before deciding to participate in the network. I’ll dive deeper into the landscape in an article next week (for those looking to dive in deeper).

Inference Capital Markets Overview

Read more

At this point, it’s pretty clear that Onchain AI is one of the most asymmetric category in Crypto. Venice has proven that a Web3 team can launch a product that’s popular outside of the CT bubble (i.e. privacy-first multi-modal inference platform designed for consumers).

Multiple Onchain/DeAI teams are racing to plant the flags, targeting different gaps, following the success of Venice.

What I’m seeing is that there are still many more undertapped, underappreciated niche that’ll highly likely drive billions of capital onchain in the next couple of years. Gonna talk about the first 2 that I believe — is the most asymmetric niche worth building for (or investing in if opportunity arise).

Let’s dig in ↓

1. Onchain agent harness

Harness is everywhere. OpenClaw started the general harness wave, Hermes continue the legacy, Codex & Claude Code dominate coding agent harness, more pop up left & right with their own differentiation.

The paper above involved a study that tests whether a difference in harness would impact cost & quality (keeping the base models constant). Turns out every model gets 33-61% cheaper.

An agent harness is essentially a harness that gets wrapped around a model or an LLM. By itself, a model can only think and talk. Harness gives the model an ability to not just talk, but do things.

We’re now in Q3/2026 and the harness is so good to the point that it’s capable of automating a lot of things for you — research, coding, teaching, designing, content, data analysis, the list is endless.

The rule of 3

Read more

I’ve been consistently using Hermes everyday for the past 4 months and he’s already a core part of my research & investment workflows.

Before Hermes, it was just me relying on Grok, Claude, Gemini for basic high-level search + all the other niche tools for in-depth search.

After Hermes, it’s calling the right skills/tools, wait for Hermes to finish researching, and then cross-check his work.

It took a while to refine the research pipeline to cover the breadth of assets/strategies I’m actively engaging with (equities, crypto, prediction markets, and more) but here are the top 3 workflows that have been the most rewarding.

1. Daily Grok Alpha Trackers + X Bookmark + Alpha Synthesis

There’s a lot of alpha on X but the algo nowadays makes it next to impossible to see your favorite accounts on your page. If you want to stay on top of everything, it’s literally impossible.

This is where Hermes comes in

Hermes can create cron jobs that automatically track your favorite analysts and deliver the summary of each of their posts to you every single day.

Hermes can also look at your bookmarks, put priority on the ones that fit with you the most, summarize, and deliver them to you.

My workflow: I’m currently tracking 11+ analysts across macro, equities, and crypto. Hermes summarizes their posts in 24+ hr timeframe, include the links to the post, and deliver them to my Discord in an actionable way (tailored to my preference/portfolio/strategies).

Same rationale also applied to bookmarks. I tend to bookmark between 5-15 bookmarks per day. Hermes categorizes them into 3 Tiers (high, medium, low importance), and deliver them everyday.

Lastly, Hermes takes analysts outputs + bookmarks + rotating external sources and synthesize them all in Top 10 alpha synthesis cron job. So instead of me going through each individual summary, I read Top 10 first and dive deeper on headlines that interest me. Everything in Top 10 gets ingested into Hindsight (external memory) and the learnings are used to improve the context for tomorrow workflows.

The rotating external sources shift from Monday to Sunday — from arxiv research papers, hedge fund letters, onchain flows, derivatives, and more. Using rotating external sources help prevent getting stuck in an echo chamber. Contrarian takes help ground the analyses, keeping me and the agent rational.

There are 3 workflows here but they’re the bread-and-butter of my Hermes research workflow.

On top of this, champion loop + feedback sweep loop (which I talked about 2 weeks ago), provide continuous feedback on daily basis, giving the agent grounds to improve the format, the content, and the analysis in each workflow.

The Requirement: Grok or X Premium subscription to use x_search for Grok Alpha Trackers workflow. X API v2 for X bookmark. Alpha Synthesis you can use any model you like (I use DeepSeek v4 Pro - cheap & provide good summary). External memory provider (I use Hindsight).

2. Onchain Forensics

If like me, you’re actively investing onchain, you’d probably often find yourself wondering if there’s going to be anybody dumping on you.

Most onchain tokens (that are investable) these days are often fair-launched or launched in a launchpad somewhere in the last 1-2 years. This makes understanding the token holders concentration, onchain buy/sell patterns, and token movements really crucial.

Investing in a project early doesn’t mean you’re going to be successful. It’s investing in a project whose team is token-aligned + whose token holders are properly distributed towards holders that are aligned.

Manually digging through onchain explorers can be difficult and time consuming.

This is where Hermes comes in.

Hermes can look at a token contract address, check who are the top holders, see if they’ve been buying or selling across any time frame, map out token holders behavior and identify whether you should continue holding, selling, or start accumulating.

My workflow: I have 3+ daily workflows that track my main onchain holdings. The agent checks large whale movements (buy/sell/transfer), flag recurring patterns, provide daily & weekly summary on changes in holders, and cross check with X sentiment. The goal is to get information from both onchain & X as earlier as possible, cross-checking them to see if both aligns.

If onchain continues to dump while X sentiment is positive, we can probably glean that something is wrong and can investigate further.

If X sentiment is negative but onchain shows continuous inflows, we could glean that somebody is probably accumulating ahead of something.

The Requirement: I’m using x402 via AgentCash & BlockRun for this workflow. Seed the wallet once and point Hermes to use tools like Nansen, BlockRun SQL, Surf, Base RPC (free). Cookie MCP is used for X sentiment.

Bonus

Been loving Exa Monitors feature. It’s a feature that basically helps you track things that matter to you. It feels like a search radar that surfaces things for you every day.

Firecrawl also has the same feature but more granular (can surface things every hour or you can set the monitor cadence yourself) but for me, the daily Exa Monitors is enough to surface meaningful news & signals from more sources beyond what I usually track.

Thanks for checking out the article, if you have any feedback or questions, feel free to drop them below

Leave a comment

Want to Continue Reading?

Check out the articles below (if you haven’t yet)

• 3 Months with Hermes - Things I wish I knew when I started

• The Hermes Sensei Loop

• Hermes Analyst 10x Better

• 6 Workflows, 6 Lessons, 60 Days

• Hermes as an Onchain Analyst

• Hermes as a Real-time Analyst

• Hermes analyst workflow essentials

• Hermes as the Ultimate Analyst — I’ve found the gist for my ultimate analyst

• 1 Month with Hermes — I’ve been using Hermes wrong all along

• Hermes, $200 and 30 skills later — here are the 3 skills that’re worth it

If I were to pick 2 things we learned this week, it’s

1. Competition between frontier labs are getting more intense

2. The moment for decentralized AI is here

SpaceXAI came out with Grok 4.5, Meta with Muse Spark 1.1, OpenAI GPT 5.6 — 3 labs duking it out in one single week.

Anthropic responded by lifting the weekly limit & extending Fable 5 usage until July 12 + there’s a possibility that they launch a cheaper/distilled version of Fable 5 to compete with GPT 5.6.

Separately, China is considering to slow down on the Chinese open-source AI wave. According to Dylan Patel (SemiAnalysis),

“There are multiple Chinese model labs who are telling all the inference guys our next model is not going to be open source, we’re going to license it to you. Open source is dying, quickly”

If the west is restricting frontier intelligence and the east is controlling it, how do the masses get access to it?

All roads lead to open-source intelligence & decentralized intelligence.

Diving deep into the state of the market + development in onchain/decentralized AI this past week + where dip-buying opportunities could be

Let’s dig in ↓

The State of the Market

Read more

Venice AI latest fundraising round of $65M at $1B “equity” valuation started a huge debate between equity <> token.

The same type of debate that we’ve seen happening time and time again after a major company acquiring other smaller projects, leaving the token holders out of the deals.

While there are hundreds of rationale supporting both sides of the equation, the best type of investment comes from being 100% aligned — be it with equity shareholders or token holders.

In this article, I’m attempting to identify the top token-aligned AI projects within the onchain AI landscape.

Let’s dig in ↓

Alignment is a Spectrum

Alignment is not a binary thing, it’s not a yes or no, it’s a continuum of business & design choices around tokenomics, monetization, incentive mechanism, and more.

A project can have a feature that aligns with token holders, another can have 5 features but might redirect value (revenue, operating profit) to equity shareholders instead.

For the sake of simplicity in our article, we’ll divide alignment into 5 tiers

• Tier 5 — Not Aligned

• Tier 4 — Trying But Emissions Win

• Tier 3 — Decent Alignment

• Tier 2 — Strong Alignment

• Tier 1 — Most Aligned

Since there are many onchain AI projects, I’ll pick the ones that are on my radar, many of which are either (i) projects which have PMF / tangible revenue / vibrant ecosystem (ii) promising projects with unique design / roadmap.

Key Criteria when judging whether a project is aligned

In order to understand whether a project is aligned or not aligned, one has to ask these questions

• Does the project generate revenue?

• If yes, where does the revenue accrue to? Equity or token? If no, what’s the plan?

• What’s the utility of the token? governance? payments? If staking, where does the yield come from? from revenue or emissions?

• If the token is used as emission or incentives, does the project have token sinks to offset the emission?

• Is the token designed as a central part of the project? If not, what’s the point of the token?

• If a business 10x’ed, what happens to the token?

And beyond these questions, it’s all the little things — what projects do when they start making money or when token start unlocking, how do they handle OTCs, how they manage the sell pressure, how they fund operations & support future growth, etc.

Alignment is a combination of all the little things, the little decisions that show that the team cares — care about themselves or their clients, or the shareholders, or the token holders, or a combination of a few of these stakeholders.

Tier 5 — Not Aligned

Most crypto projects fall under this category. They raised $$$ from VCs during the last cycle. Launched a governance token or L1/L2 token because they had to, because they treated token as only capital bootstrapping mechanism.

Why?

Because they could.

Capital was everywhere. If a founder had a modicum of knowledge on a certain vertical, he’d be able to raise from VCs or the community.

This led to a horde of useless tokens supporting useless projects. The most popular one that tend to be at the brunt of it is

Read more