General Intelligence Capital released ATLAS, a framework for autonomous AI trading agents that improve through market feedback by applying Andrej Karpathy's autoresearch methodology to financial markets. The GitHub repository gained 737 stars within three days of its March 11, 2026 release, demonstrating measurable agent performance improvements across 378 trading days from September 2024 to March 2026.
Autoresearch Methodology Treats Sharpe Ratio as Loss Function
ATLAS applies Karpathy's autoresearch paradigm to trading, where agent prompts function as weights and Sharpe ratio serves as the loss function. The system operates through evolutionary optimization: if performance improves, the git commit survives; if not, git revert. This approach enables agents to autonomously refine their own decision-making strategies based on real market outcomes.
The system architecture features 25 agents debating markets daily across 4 layers, with every recommendation scored against actual outcomes. The worst-performing agent automatically gets its prompt rewritten, creating a continuous improvement cycle driven by market feedback rather than manual tuning.
Measurable Performance Improvements Across Multiple Asset Classes
ATLAS demonstrates quantifiable learning through iterative prompt refinement:
- Financials agent: Sharpe ratio improved from -4.14 to 0.45
- Emerging markets: Sharpe ratio improved from -0.45 to -0.06
- Semiconductor: Sharpe ratio improved from -0.26 to -0.06
These improvements resulted from 378 days of evolutionary optimization against 18 months of real market data, with trained prompts representing the cumulative learning product.
Multi-Agent Coordination Integrates Structured Market Information
The accompanying academic paper "ATLAS: Adaptive Trading with LLM AgentS Through Dynamic Prompt Optimization and Multi-Agent Coordination" (arXiv:2510.15949) describes ATLAS as a unified multi-agent framework integrating structured information from markets, news, and corporate fundamentals. Agents provide daily market recommendations across different sectors and asset classes, with recommendations scored against actual market performance to guide the rewriting process.
Framework Open Sourced but Trained Prompts Remain Proprietary
General Intelligence Capital released the ATLAS framework and results documentation on GitHub, but the trained prompts themselves remain proprietary. The repository contains the architectural approach and methodology, allowing researchers to replicate the autoresearch process, but not the specific optimized prompts resulting from the 378-day training period.
Key Takeaways
- ATLAS applies Andrej Karpathy's autoresearch methodology to financial markets, treating Sharpe ratio as the loss function for agent improvement
- The system achieved measurable performance gains across 378 trading days, with the Financials agent improving Sharpe ratio from -4.14 to 0.45
- Architecture features 25 agents debating markets across 4 layers, with the worst performer automatically rewritten based on market outcomes
- Framework demonstrates that LLM agents can autonomously improve through evolutionary prompt optimization using clear market feedback signals
- GitHub repository gained 737 stars within 3 days, though trained prompts remain proprietary to General Intelligence Capital