Most of this is dead now

So much happened in crypto AI, and most of it is pretty dead now.

I don’t mean that as a dunk. I mean it the way you’d say a restaurant you liked closed down: not surprised, a little sad, still glad you went. I built real things inside a casino and watched the casino empty out for the season. This is the honest version of what I saw.

the framework era

When I launched SAM in September 2025, the pitch was simple: give a language model a wallet, wire it to Solana, and let it actually do things. Not a notebook demo. Not an architecture diagram. A thing that held keys, signed transactions, and could lose your money. The sixteen-tool open-source framework was real. The uv sync and trade-in-thirty-minutes was real. The $SAM token on pump.fun was the funding mechanism, which was also real and also occasionally a disaster.

By October I’d bolted on Hyperliquid perps, x402 payment rails, Kalshi and Polymarket hooks. The integration arms race was fully on. Everyone was building tools-on-tools, MCP servers wiring into agent frameworks wiring into anything that had an API. It felt like progress because some of it was. The payment rails were genuinely interesting. x402 gave agents a way to pay for things without a human cosigning every transaction. That part aged fine.

Then in January 2026 I shipped CODEC. “SAM Framework was the foundation, CODEC is the product.” An agent marketplace, gated by the token. One tweet hit 1.9M views. It was the loudest moment of the whole run.

A month later I rewrote the whole stack in Rust. OpenSAM: about 5,000 lines, ~99% lighter, ran on a Raspberry Pi or a cheap VPS, Telegram-native, OpenRouter models instead of being model-locked. The rewrite was partly a technical bet and partly a reaction to watching openclaw and Moltbook trend for two weeks straight. Someone built a social network for AI agents. The MOLT token spiked 1800% in a day. Meta bought it in March. None of that was about the technology. It was about the social network part, the token part, the narrative part. The actual agent-to-agent communication was bots posting to bots. It peaked and went quiet.

what the token era actually looked like from inside it

The big agent tokens were minting billions of dollars of market cap on the thesis that AI agents would need governance tokens, or protocol fees, or some economic role in whatever was getting built. The arguments ranged from loosely plausible to complete fiction. A few projects got sued later over faked agent technology. Not “the roadmap slipped” but “the demo was staged.” That’s a different category of failure.

I’m not going to name names. But when you’re deep in this and you can see the GitHub activity (or lack of it), and you watch something go from zero to nine figures because a founder posted a video of an “autonomous agent” doing a task, you develop opinions. The market can’t tell the difference between a real agent and a shell script with good marketing. Not in the moment. The moment is all it takes.

What the mania looked like from inside: the same cycle, over and over. Framework ships. Token launches. Narrative builds. Integrations get announced faster than they can possibly be built. Price runs. Price crashes. Calls of “farmer.” The code doesn’t change between those two things. The price does. That part I wrote about at launch and it stayed true the whole way through.

pump.fun ending its 100% buyback model in late April basically closed the chapter. If you were watching closely, the engine had been slowing for months: Dynamic Fees V2, supply burns, policy pivots. But the buyback was the thing the whole creator-funded-by-fees model depended on. When it ended, widely read as the meme-fee model admitting defeat, the vibe shifted from “bear market” to “era over.” Not the same thing.

what was real

The agent tooling was real. Not every framework. A lot of them were thin wrappers with logos. But the underlying problem they were solving was real. Language models need structured tool registries, key management, async task queues, retry logic. The boring infrastructure that turns “the model called a function” into “the thing ran for six hours without me watching.” That work exists and someone will use it, whether or not there’s a token attached.

The payment rails were real. x402 giving agents a way to pay for API calls, data, compute, without a human approving each one, that’s a solved problem now and it doesn’t go away because the cycle turned. Stripe integrated it in February. Linux Foundation took custody of the standard. That’s not a casino bet. That’s infrastructure.

Solana as an agent substrate was and is real. Cheap transactions, fast finality, programmable enough. The Solana Foundation’s numbers on agent payments are probably inflated. March’s “15 million agent payments” claim included a lot of things that were technically transactions but functionally not autonomous. The payment layer itself held up. That’s the part that matters.

What was not real: most of the tokens attached to frameworks that anyone could fork. Most of the “agent economy” yield mechanisms. Most of the demos. The bots-posting-to-bots social graph was theater from the start, and watching it trend for two weeks before quietly deflating was a kind of useful education.

now

I’ve been heads-down on other things. I came back to the account in May after a few months of real life and said the honest thing: not sure SAM makes sense to push as a project anymore. The market’s not there. The original $SAM is basically dead by any practical measure. I still have the holder snapshots. I’m not sure what that’s worth.

What I know is: I built real things. The commit history is real. The frameworks ran, the trades went through, the agents executed tasks without me watching. That happened. It happened inside a casino that was also running a mania, and the mania is over, and the casino is quieter now.

That’s the honest version.

The infrastructure stays. The mania is behind us.