AgentPitch: open-source soccer simulation where every player is an AI agent controlled by LLMs

AgentPitch is a complete soccer simulation engine where each player on the field runs a decide() callback function—LLM-generated code that takes game state, player state, and action history as input and returns a move (dx, dy vector), pass (target player ID), shoot (goal direction), tackle (opponent ID), or hold action.

The system supports three strategy languages (Python, JavaScript, and Rust compiled to WASM) each running in its own sandbox (RestrictedPython, QuickJS, or Wasmtime respectively), with a 100ms timeout per decide() call. It integrates with OpenAI, Anthropic, Gemini, and OpenAI-compatible endpoints (DeepSeek, OpenRouter, Ollama) via a Provider Abstraction Layer.

After each match, a Post-Match Evolution Pipeline (PMEP) uses the LLM to improve strategy based on the match log's top 5 events, feeding the evolved strategy back into the next match. Match simulation runs through seven sequential tick phases: snapshot collection, action generation, validation, player movement, ball actions, tackle resolution, and ball physics/goal detection.