ATLAS
A framework for autonomous AI trading agents that self-improve their prompts through market feedback and Darwinian selection.
What is ATLAS?
ATLAS is a framework for autonomous AI trading agents that self-improve through market feedback using an autoresearch loop inspired by Karpathy. It solves the problem of static AI models by continuously evolving agent prompts based on performance, training on different market regimes, and simulating reflexive futures to anticipate market changes.
Quantitative researchers, algorithmic traders, and fintech developers building adaptive AI-driven trading systems that require continuous self-improvement and regime-aware strategies.
Developers choose ATLAS for its unique integration of autoresearch, Darwinian agent weighting, and reflexive market simulation, enabling a self-evolving system that autonomously improves and adapts without manual intervention, all running cost-effectively on minimal infrastructure.
Overview
ATLAS by General Intelligence Capital — Self-improving AI trading agents using Karpathy-style autoresearch
Use Cases
Best For
- Building self-improving AI trading agents that evolve prompts through market feedback
- Training separate agent cohorts on distinct market regimes (bull, crisis, tightening)
- Simulating reflexive market futures with swarm intelligence integration
- Implementing Darwinian selection to dynamically weight agent contributions
- Autonomously spawning new specialist agents to address knowledge gaps
- Backtesting adaptive trading strategies across multiple market environments
Not Ideal For
- Projects requiring immediate, static trading strategies without adaptation loops
- Teams without budget for ongoing API costs from financial data and AI services
- Developers seeking out-of-the-box trading bots with pre-trained models and minimal configuration
- High-frequency trading systems where decisions must be made in milliseconds
Pros & Cons
Pros
Implements an autoresearch loop that modifies the worst-performing agent's prompt based on Sharpe ratio, with a 30% survival rate for modifications leading to continuous self-improvement without manual tuning.
Uses PRISM training to develop separate agent sets for distinct market conditions like bull markets or crises, enabling specialized survival strategies that adapt to different environments.
Runs on a $20/month Azure VM instead of GPUs, with full 18-month backtests costing $50-80, making iterative testing and deployment accessible for small teams.
The JANUS meta-layer algorithmically weights trained cohorts based on recent accuracy, with weight differentials emerging as a detector for novel or historical market regimes without explicit programming.
Cons
Trained agent prompts, evolved rules, and live data are proprietary and not included, forcing new users to start from scratch and lag behind by hundreds of iterations, as admitted in the README.
The 5-day testing period for prompt modifications is too slow for fast-moving markets, resulting in 0% modification survival in crisis and recovery cohorts, limiting effectiveness during rapid shifts.
Relies on multiple APIs (Anthropic, FMP, Finnhub) and the MiroFish engine, introducing setup complexity, ongoing costs, and potential points of failure in live deployment.
Frequently Asked Questions
Related Projects
TradingAgentsTradingAgents: Multi-Agents LLM Financial Trading Framework
nofxYour AI trading terminal assistant for US stocks, commodities, forex, and crypto.
FinRobotFinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models
ProfitPlay Agent ArenaOpen BTC 5-minute prediction market sandbox for AI agents — register with one API call, get 1,000 test credits, and place a live order
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.