ATLAS is a framework for developing self-improving AI trading agents that leverage Karpathy-style autoresearch principles. It targets technically savvy users interested in autonomous systems for financial markets, offering a novel approach to prompt optimization driven by market feedback and evolutionary selection. The system aims to autonomously refine trading strategies, potentially leading to enhanced performance.

How It Works

The core mechanism applies Karpathy's autoresearch loop to financial markets, where agent prompts function as optimizable weights. A multi-layered architecture (Macro, Sector, Superinvestors, Decision) facilitates daily market analysis and debate among 25 agents. Performance is evaluated via Sharpe ratio against real market outcomes; the worst-performing agent's prompt is iteratively modified and either committed (if performance improves) or reverted. This Darwinian process, running on a low-cost VM, optimizes agent prompts over time, with agent trust quantified by dynamic weights.

Quick Start & Requirements

• Primary install / run command: Framework architecture and pipeline structure are provided. Specific setup details beyond API integrations are not detailed.
• Non-default prerequisites and dependencies: Requires API access to Claude Sonnet (Anthropic), data providers (FMP, Finnhub, Polygon, FRED), and an Azure VM ($20/month). No GPU is required.
• Estimated setup time or resource footprint: Backtesting for 18 months cost approximately $50-80. Live deployment utilizes a low-cost VM.
• Links: generalintelligencecapital.com, chris@generalintelligencecapital.com.

Highlighted Details

• Backtest Performance: 18-month backtest (Sep 2024 - Mar 2026) yielded +22% in the deployment phase (173 days), with a best individual pick achieving +128%.
• Autoresearch Efficacy: Out of 54 prompt modifications attempted, 16 (30%) survived, indicating a challenging optimization landscape.
• Agent Improvement: Significant Sharpe ratio gains were observed for specific agents, e.g., Financials improved from -4.14 to 0.45.
• Key Insight: The orchestration and decision-making layer is identified as a critical bottleneck, potentially more impactful than individual agent intelligence improvements.

Maintenance & Community

The project is led by Chris Worsey (CEO & Technical Founder). Further details on community engagement, roadmap, or ongoing maintenance are not provided in the README.

Licensing & Compatibility

• License type and notable restrictions: The repository's license is not specified in the README.
• Compatibility notes for commercial use or closed-source linking: Designed for financial market analysis using specific third-party APIs and cloud infrastructure.

Limitations & Caveats

The core intellectual property—trained agent prompts—is proprietary and not included. Key components like active management rules, agent scorecard data, and live deployment infrastructure are also absent. The high rate of prompt reversion (70%) suggests the autoresearch process requires significant tuning and may not consistently yield improvements. The absence of a specified license poses a significant adoption blocker.