ATLAS
Open-source self-improving multi-agent trading framework — 25 agents across four layers, with Darwinian prompt evolution scored by Sharpe ratio.
ATLAS is an open-source multi-agent trading framework from General Intelligence Capital that applies Karpathy-style autoresearch to markets. Twenty-five agents debate across four layers — macro, sector desks, superinvestor philosophy, and decision — with agent prompts treated as weights in a Darwinian loop scored by rolling Sharpe ratio. The worst performer gets its prompt rewritten; the change survives only if performance improves.
The system spawns new specialist agents when it detects knowledge gaps and trades live. The architectural approach and methodology are public; the trained prompts remain proprietary. Runs on a ~$20/month VM with no GPU.
