autocontext

Why we included this project

Teams that rely on coding or support agents often watch them repeat the same mistakes run after run. autocontext addresses that by treating every task as a chance to learn: you give it a goal in plain language, it runs the task against an evaluation, keeps what worked, drops dead ends, and leaves behind traces, reports, playbooks, and datasets for the next iteration. Because everything is stored as files, you can inspect, diff, replay, or feed that accumulated knowledge into training, which makes it practical for distilling stable behavior into a cheaper runtime. It works as a Python CLI and library, a TypeScript/Node package, and can drive Claude, Codex, Pi, or any OpenAI-compatible endpoint, including self-hosted models on vLLM or Ollama. That combination of a learning loop and flexible deployment makes it a solid pick for teams that want their agents to get measurably better over time.