Agent Apprenticeship
An open ecosystem where AI agents complete real-world tasks through workflow loops, get evaluated by mentor agents or humans, and turn the results into reusable training data.
Agent Apprenticeship is infrastructure for agents to complete tasks through iterative workflow loops, improve through repeated execution, and be evaluated by mentor agents, users or human experts, turning finished work into reusable experience and training signals for future agents. It is designed for a compounding cycle: economically valuable task execution generates learning signals, those signals improve future work, and future work creates new experience for the ecosystem.
Its seed dataset includes more than five hundred curated real-world tasks, nearly five hundred reusable agent lessons, over a thousand full execution traces and task episodes, more than five hundred work-experience compilations, and tens of thousands of structured experience records. It works with local agents including Codex, Cursor, Claude Code, OpenClaw, OpenCode and Hermes Agent across different model providers, and installs with a single npx command. It is written in Python and licensed under MIT.