Agent Apprenticeship Guide

Key takeaways#

Agent Apprenticeship is an open resource and CLI ecosystem for collecting real-world agent work experience, mentor feedback, workflow loops, and reusable training signals. The repository describes a living system where apprentice agents complete tasks, learn from mentor agents or humans, and convert completed work into future agent experience.

What it provides#

The project includes an npm-based setup path through npx agent-apprenticeship init or a global agent-apprenticeship install. The README describes apprenticeship modes, mentor model provider configuration, local settings, a seed dataset, reusable lessons, execution traces, task rollouts, and structured experience records. It is aimed at agents that operate over longer workflows rather than single-turn chat.

The concept is useful for builders because most agent systems lose learning signal after a task finishes. Agent Apprenticeship frames each task as part of a feedback loop: plan, execute, review, record experience, and improve the next run. The repository mentions use with local agents and coding environments such as Codex, Cursor, Claude Code, OpenClaw, OpenCode, Hermes Agent, and custom agents.

Setup outline#

Install with npx agent-apprenticeship init, use defaults for a first run, then inspect settings with apprentice settings and apprentice doctor. Configure the apprentice agent, mentor provider, and apprenticeship mode. Provider keys can be stored in ~/.agent-apprenticeship/.env.local or supplied through the shell environment.

How teams can use it#

A solo builder can use the project as a workflow journal for agents: record what happened, keep reusable lessons, and make future runs less repetitive. A research team can use it as a structure for collecting task rollouts and mentor feedback. An organization can adapt the idea into evaluation and training pipelines where completed work creates structured data for later agents.

What to watch#

Do not assume the seed dataset or workflow loop automatically improves your agents. Review the data your agents create, remove sensitive material, and decide which lessons are safe to reuse. If you connect paid model providers, set spending limits before running long autonomous workflows. Treat mentor feedback as a quality layer, not as a guarantee that every resulting lesson is correct.

Best fit#

Agent Apprenticeship is best for builders experimenting with long-horizon agents, reusable agent experience, and human or mentor-in-the-loop evaluation. It is a resource for designing agent learning loops and a CLI for getting started, not a generic chatbot framework.