a recursive self-improving harness designed to help your agents (and future iterations of those agents) succeed on any task
Autocontext is a recursive self-improving harness designed to help LLM agents iteratively refine strategies for a wide variety of tasks. The system runs a multi-agent loop where agents collaboratively evolve and score solutions through tournament matches for game scenarios or LLM judge evaluation for agent task scenarios, with Elo-based progression gating. It's built in Python and TypeScript, with the Python control plane (`autocontext` package) handling scenario runs, simulations, investigations, persistent knowledge management, and MCP/HTTP surfaces for external agents, while a Node/TypeScript toolkit provides CLI tools and an interactive terminal UI. The project supports multiple LLM providers including Anthropic's Claude, OpenAI-compatible endpoints, local models via Ollama, and command-line tools like Pi and Codex, and includes templates for common optimization tasks like content generation, prompt optimization, and RAG accuracy tuning.