Running a nine-agent hedge fund with GPT-6 Astra
Argus uses nine agents to scan crypto markets 24/7, with GPT-6 Astra writing strategy code, and only strategies that clear hard out-of-sample gates on Sharpe and drawdown get promoted; it is paper trading only and runs offline.
I let GPT-6 Astra run a hedge fund. Argus is a hedge fund in code. 9 agents scan crypto 24/7. GPT-6 Astra writes candidate strategies. Only the ones that clear hard out-of-sample gates get promoted. Paper trading only. Runs offline. No API key. The goal was simple - generate ideas overnight, kill almost all of them, and only keep a strategy if the statistics survive data it never saw. Every agent exists because of a failure mode I hit earlier: 1) Watcher - OHLCV, L2 book, perp funding. Stat-arb, vol, factor residual, order-flow. Nothing else. 2) Hypothesis - Astra writes real Python under a contract. Not a take. Not a prompt dump. Code. 3) Backtester - vectorized, next-bar only, fees + slippage. Look-ahead dies here. 4) Validator - grades a sealed strategy on an out-of-sample split it did not fit on. Sharpe above 1.5, max drawdown under 15%, hit rate above 55%, t-stat above 2.0. Miss one gate and it is dead. 5) Checker - hidden vault split no other agent has seen. Look-ahead test. Code review. This is the desk that exists so Hypothesis cannot grade its own homework. 6) Risk - 1/4 Kelly, exposure caps, 1x leverage. Full Kelly is how you go broke while being right. 7) Execution - paper broker stub. Live trading is disabled on purpose. 8) Scribe - report.md + events.jsonl. Optional Telegram. The run is reproducible. 9) Orchestrator - nightly loop. I do not touch config at 3am. One rule from day one: > no agent grades its own output > the vault is data the writer never touched > a kill switch that trusts the bot's own claim is not a kill switch Most nights it promotes nothing.
This post quotes another post: “Long read: building hedge-fund-grade trading models with GPT-6 Astra”