ATLAS
Self-improving trading agents whose prompts are rewritten when they lose, from General Intelligence Capital
By General Intelligence Capital
Live GitHub metrics
Live GitHub data · Oct 11, 2026No tagged releases. Repository created Mar 11, 2026.
What it is
ATLAS runs 25+ agents in four layers (macro, sector desks, investor personas and a decision layer with an adversarial risk officer) that debate markets daily. Every recommendation is scored against outcomes, the worst agent's prompt is rewritten and the change is kept only if performance improves, using git commits and reverts as the training loop in the style of Karpathy's autoresearch. The README states it runs live with real capital; treat that as the maintainers' claim.
Custom licence that is not OSI-approved; read it before reuse.
In the maintainers' words: “ATLAS by General Intelligence Capital — Self-improving AI trading agents using Karpathy-style autoresearch”
Markets
- US equities
Tags
- Self-improving
- Multi-agent
- Prompt evolution
- Debate