forsy-ai/agent-apprenticeship

Agent Apprenticeship is a system that lets AI agents get better at real work by actually doing tasks, getting feedback, and turning that experience into reusable knowledge. Think of it as an apprenticeship program for AI: an agent attempts a task, a mentor (either another AI or a human) reviews the work, and the lessons learned get fed back into a shared ecosystem so future agents start with more know-how. At a high level, you give the tool a task, something like "create a market map for AI procurement tools." An apprentice agent works through it in iterative loops, refining as it goes. When the run finishes, the system produces an "experience compilation" that captures what happened. You can then install that compilation as "runtime training," meaning the next time an agent runs a similar task, it benefits from prior experience. The project ships with a seed dataset of over 500 curated tasks, nearly 500 reusable lessons, and thousands of execution traces and work episodes to get things started. The tool is for people already using AI coding or task agents, Codex, Cursor, Claude Code, and others, who want those agents to improve over time instead of starting fresh each run. A founder might use it to train an agent on producing consistent competitive analyses, a PM might use it to build up a reliable release-checklist workflow. You can run fully autonomously, in expert-led mode, or in a custom organizational setup, and you choose whether your experience compilations stay private or contribute to the public ecosystem. What's notable is the compounding loop: every completed task generates structured learning signals, and those signals become training material for the next round. The project is built to work across model providers (OpenAI, Anthropic, Google, OpenRouter) and supports custom agent commands, so it's not locked to one toolchain. Setup is a single command, and configuration is handled through a local settings file.