ATLAS is a multi-agent AI trading framework that orchestrates 25+ LLM-based agents across four layers (Macro, Sector, Superinvestor, Decision) to generate and execute trading recommendations. It implements an autoresearch loop that treats agent prompts as trainable weights, using Sharpe ratio as the loss function to iteratively improve performance through git-based keep/revert cycles. The system integrates MiroFish swarm simulation to generate synthetic future scenarios and forward-train agents, and includes a JANUS meta-layer that dynamically weights recommendations from multiple trained cohorts to detect market regimes. The codebase demonstrates strong architectural separation of concerns, with reusable components for meta-learning, scenario generation, and agent training, though some modules depend on proprietary data clients and configuration not included in the repository.