agent-apprenticeship

Why we included this project

Agent Apprenticeship sits on top of the coding agents you already run, like Codex, Cursor, or Claude Code, and turns each finished task into something the next run can learn from. A mentor model or a human reviews the output, and the whole session gets packaged into an experience compilation you can load as training for later work. So instead of every run starting cold, prior runs leave behind lessons and traces that shape the next attempt. The bundled seed dataset, with hundreds of real-world tasks and thousands of execution traces, shows the format and gives you a base to contribute your own signals. If you want agents that improve over time rather than just answer one-off prompts, this is a self-contained way to build that feedback loop locally.