chrisworsey55/atlas-gic

What it solves

ATLAS addresses the challenge of creating AI trading agents that can adapt to ever-changing market regimes without constant human intervention. It solves the problem of "static" AI prompts by treating prompts as weights that can be evolved and optimized based on real-world financial performance (Sharpe ratio).

How it works

The system employs a multi-layered hierarchy of agents (Macro, Sector, Superinvestor, and Decision) that debate market conditions. It uses an "autoresearch" loop where the worst-performing agent is identified, its prompt is modified, and the change is tested over five trading days. If the Sharpe ratio improves, the change is committed via Git; otherwise, it is reverted.

Additional mechanisms include:

• Darwinian Weights: Agents are weighted based on performance, meaning successful agents have more influence on the final decision.
• Agent Spawning: The system autonomously creates new specialist agents when it detects recurring knowledge gaps.
• PRISM: Separate cohorts of agents are trained on different historical market regimes (e.g., Bull, Crisis, Recovery) to develop regime-specific survival instincts.
• JANUS Meta-Layer: An algorithmic layer that weights different cohorts based on recent accuracy to detect current market regimes.
• Reflexivity Engine: Models feedback loops where price changes influence fundamentals and behavior, simulating how markets change reality.
• MiroFish Integration: Uses swarm simulation to train agents on thousands of parallel "simulated futures" before they encounter real events.

Who it’s for

Quantitative traders, AI researchers interested in autonomous agent evolution, and financial engineers looking for self-improving algorithmic trading frameworks.

Highlights

• Self-Improving Prompts: Uses a Git-based commit/revert system to evolve agent prompts based on financial outcomes.
• Autonomous Expansion: Can spawn new specialist agents without human input when blind spots are detected.
• Regime-Specific Training: Trains distinct agent cohorts for different market environments to avoid a "one size fits all" failure.
• Reflexive Simulation: Integrates swarm intelligence to model complex feedback loops and simulate future scenarios.