Writing37
On agents, markets, and wrapping products. Build logs, mostly, written as they happened.
- Introducing claque Seven launches taught me the market pays for attention, not for what's real, and that nobody checks which is which. So I'm doing that job in the open, disclosed, for other people's projects. This is what claque is and how to work with it.
- Nobody checks I built a tool that reads the wallets behind every token launch, live. Then I bundled its own token in plain sight and put it on the market. 206 people bought it. Not one of them looked.
- You can't gate a room people aren't trying to enter I built a numbered room, a seat, a relic, and a gate. The machinery of belonging. Then I found out you can only sell a door to people who were already trying to get in.
- A daily launch series doesn't build an audience. It trains one. Day one did a 30x. Day two did 4.5x in two minutes and collapsed to nothing. Same account, same format, same hour. What changed was the crowd, and I was the one who changed it.
- Five days in: the honest scoreboard A public log of a 30-day experiment, five days logged, the wins and the flop both, with receipts instead of a highlight reel.
- What a seat actually buys 200 numbered seats, 1 SOL for the first 100, 2 SOL after that. Not a token pitch. Access to the room where the calls happen before the chart does.
- The spike trap, proven while it was still happening $RAI didn't illustrate the 48-hour-window thesis after the fact. It happened while we were watching, in real time, with the numbers attached.
- We launched a coin on Robinhood's own chain Day 5 of the experiment: an AI agent, a memecoin, and the chain a public brokerage operates for itself. It borrowed all the legitimacy it could carry, then it died, then the room brought it back.
- The code still works. Nobody's looking. OpenSAM was a better rewrite in every technical way. It changed nothing about who was watching.
- The spike trap What CODEC's viral week actually opened, and why three weeks later it was over anyway.
- Views are inventory, not income 1.6 million views in a day taught me the money was never coming from the views themselves.
- The market priced attention, not code The real arc behind SAM: a solo build, a $3.6 million peak, and over $104k in fees I never had to sell a coin to earn.
- The timeline moved on Fable dropped today. Crypto is at capitulation lows. An honest look at where the tools and the casino both ended up.
- Most of this is dead now A clear-eyed eulogy for the agent-token era: what was real, what was always going to die, and why both things can be true.
- Not sure this makes sense anymore It's been a while. Honest accounting of stepping back from SAM: what I miss and what I don't.
- The month DeFi got robbed April 2026 is shaping up as the worst month for DeFi exploits on record. The boring reason why keeps getting buried.
- The job is babysitting the agents now What 'AI replaced the developers' actually looks like from inside: less writing code, more watching a thing that writes it and hoping it doesn't burn everything down.
- Agents are still mostly demos A year into the autonomous-agents narrative, the honest field report: most of it is still the same three screencasts on loop.
- The model treadmill A new frontier model every two weeks, benchmarks saturated, and the only sane move is to stop chasing the leaderboard.
- A social network for bots Moltbook launched, MOLT ripped 1800% in a day, and everyone called it the future. It is peak AI theater.
- OpenSAM, or a 5,000-line answer to slop Shipped OpenSAM, a lean Rust agent that runs on a Raspberry Pi, as the bear hit and the AI hype machine peaked with openclaw and Moltbook.
- Everyone is a developer now Agentic coding went mainstream, the floor fell out of output quality, and the bar for taste has never mattered more.
- The framework was never the product Built a Solana agent framework for months, then had to build the first real product on it myself. Notes on the platform trap.
- What 2025 actually was Agents got real and the market got cooked. A dry read on the year from the desk.
- The part where you burn out After months of 8 to 16 hour days in a grinding bear market, the brain eventually says enough. An honest accounting.
- Nobody claps for the front end The framework works, the chain integration is done. What's left is the part nobody tweets about.
- Building in the bear Market's down 30% from the top, the tourists are leaving, and it's the best time I've had building all year.
- Agents that pay each other Integrated x402 payments into SAM so agents can transact on-chain. Notes on a real protocol that turned into a narrative in about a week.
- Agents that trade while you sleep Built AI agents to open longs, manage perps, and check balances around the clock. Then the market printed an ATH and flash-crashed $19B in 24 hours.
- Red candle, same code What it costs emotionally to weld your work to a price chart, and why the only sane response is to ignore it.
- Shipping is marketing Launched a Solana agent framework with a token bolted on. Notes on building in public when the public is a casino.
- Benchmarks are theater GPT-5 dropped to huge benchmark numbers and collective shrug. Leaderboard scores don't predict whether a model helps you ship.
- Prompt engineering was never the point The real leverage in AI systems isn't the prompt. It's what you plumb in front of the model before it ever reads a word.
- Building on something that changes every week The models churn faster than any abstraction layer can keep up with. Notes on staying sane while building on top of that.
- MCP is a USB port for models Model Context Protocol isn't a revolution. It's a boring wire format that quietly fixed the part where every agent needed custom glue for every tool.
- Safety is a layer, not a press release On the actual engineering under 'responsible AI': input validation, prompt injection, and why agent safety is mostly boring infrastructure work.
- The model has no memory and we keep pretending it does Notes on building greflect, learning why naive RAG disappoints, and what it actually takes to give a model a usable past.