Agents that trade while you sleep
The pitch was clean. Wire an AI agent into a perps exchange, give it tools to open longs and shorts, check balances, pull trade history, and let it run. No sleep. No emotion. No missing a move at 3am because you were unconscious. You ship the agent, you go to bed, the agent trades. This is the thing everyone in the AI-DeFi corner has been gesturing at: autonomous execution in a market that never closes.
I spent the last two weeks actually building it.
what got wired
The Hyperliquid integration landed at the start of October. Manage trades, open positions, check balances, all wrapped as tools the agent can call. Before that, in late September, I’d done the initial groundwork for Aster, which came online first: natural language perp trading, long/short execution, trade history. The idea being that an agent could scan the market, decide on a position, and place it, all without a human in the loop.
Multi-agent support came in the same commit batch. The framework already had a tool registry and a keyring. Multi-agent meant you could now route tasks between specialized agents: one watching price action, one managing risk, one executing. The whole setup starts to look less like a single bot and more like a small autonomous trading desk that happens to live on your server.
The obvious use case is what I was pitching to anyone paying attention: let the agent trade perps 24/7 while you sleep. You set the parameters. It operates.
what the market said about that
October 6th, BTC hit an all-time high. Around $126K. The market had been constructively bullish through the whole autumn, and the ATH felt like confirmation. Crypto twitter was extremely pleased with itself.
October 7th, the next day, roughly $19 billion in positions got liquidated in a single session. ATH to bloodbath in under 24 hours.
I’m not going to tell you what caused it, because the cause doesn’t really matter for this observation. What matters is the shape of it: the market printed the highest number it had ever printed, then immediately reminded every participant that it does not care about your thesis, your agent, your risk parameters, or your sleep schedule. The automation cuts both ways. An agent that trades while you sleep will also trade into the flash crash while you sleep. It will execute your parameters into a market that has, without warning, decided to go sideways and take everyone with it.
the irony is load-bearing
There’s something genuinely funny about spending two weeks building tireless machines to trade a casino, and then watching the casino do that. Funny as in the universe made the joke so obvious that you’d feel stupid not acknowledging it.
The pitch for 24/7 automated execution has always assumed that the edge is human limitation: we sleep, we hesitate, we miss the 3am move. That’s true. But the market doesn’t exploit only that limitation. It exploits every edge it can find. The agent doesn’t hesitate. It also doesn’t have judgment. It will hold into the crash exactly as confidently as it held into the ATH, because it can’t tell the difference until the difference is already priced in.
This is not an argument against building the thing. The tools are real. The execution layer is real. Multi-agent coordination for trading is genuinely interesting infrastructure. I built it because it’s useful, and I’ll keep building it because the framework exists to make these things possible without starting from scratch.
But the October 7 session is a good calibration for the pitch. “Agents that trade while you sleep” is technically accurate. It’s also a description of a thing that was long into the correction while you were dreaming.
The machine doesn’t need you awake. The casino doesn’t need you asleep. Both of those things are equally true and always have been. I was just too busy shipping the former to think much about the latter.
Still here. Back to the frontend.