autocontext

autocontext is an open-source recursive self-improving harness for LLM agents that closes the loop between execution, evaluation, and knowledge accumulation. Instead of starting every agent run cold, autocontext persists what worked — traces, playbooks, artifacts, and distilled models — so each subsequent run builds on validated prior success. It supports Python and TypeScript surfaces, multiple LLM providers, and a structured multi-agent loop with roles for proposing, analyzing, coaching, and curating knowledge.

• Scenario Families: 11 reusable scenario families (game, agent_task, simulation, investigation, workflow, negotiation, coordination, and more) executable in both Python and TypeScript.
• Multi-Agent Loop: A structured internal loop with competitor, analyst, coach, architect, and curator roles that propose, evaluate, and gate knowledge persistence.
• Persistent Knowledge: Validated playbooks, hints, tools, reports, and progress snapshots accumulate across runs rather than being discarded.
• Multiple Surfaces: Access via CLI (autoctx), REST API server, MCP server, TypeScript/TUI operator surfaces, and external agent integration.
• Provider Routing: Supports Anthropic, OpenAI-compatible endpoints, Gemini, Mistral, Groq, OpenRouter, Azure OpenAI, MLX (Apple Silicon), Pi, and deterministic testing backends.
• Frontier-to-Local Distillation: Export stable training data and distill it into cheaper local runtimes using MLX on Apple Silicon.
• Replay and Analysis: Inspect, compare, and replay runs, simulations, investigations, and missions to understand regressions and stable wins.
• Notification Hooks: Route notifications via Slack, HTTP webhooks, stdout, or composite routing using AUTOCONTEXT_NOTIFY_* env vars.
• Quick Start: Install via pip install autocontext or npm install autoctx, then run autoctx solve --description "your task" --gens 3 to hand the harness a plain-language task and let it iterate.