Agent Apprenticeship
Turn local agent task runs into reusable experience and runtime training.
- Source repo
- ray-r-ren/agent-apprenticeship
- Stars
- ★ 1.6k
- Last updated
- 18d ago
- License
- MIT
- Primary language
- Python
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer using Codex locally wants to run a multi-step research or release-checklist task and reuse its resulting experience in later runs.
- Not a fit if
- Teams wanting a one-click hosted service instead of a local shell
- Teams unwilling to share task data or experience compilations with the public ecosystem
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
Agent Apprenticeship is a local command-line workflow-loop system in which apprentice agents execute tasks with evaluation from mentor models, users, or human experts. It supports Codex, Cursor, Claude Code, OpenClaw, OpenCode, Hermes Agent, and custom command templates, and it detects installed CLIs. Completed runs print a local run folder and an Experience Compilation path; compilations can be inspected, checked, exported, searched, pulled, or installed as Runtime Training for later work. The project also documents Public Ecosystem and Private Internal Only contribution modes, plus a Seed Dataset v0.2 containing tasks, traces, and experience records.
apprentice run "<task>" starts a task workflow using the configured Apprentice Agent. apprentice configure and apprentice configure model configure the agent, mentor model provider, and Apprenticeship Mode; custom agents can be connected with apprentice configure agent custom --command-template "...". After a run, apprentice watch <run_id> follows progress, while apprentice ecosystem inspect <experience_compilation_path> and apprentice bundle check <experience_compilation_path> inspect the resulting Experience Compilation. apprentice ecosystem export ... --full exports the full compilation, and apprentice learn install ... installs prior experience as Runtime Training; public-mode commands also include apprentice ecosystem search, inspect, and pull.
- A developer using Codex locally wants to run a multi-step research or release-checklist task and reuse its resulting experience in later runs.
- An individual who has Claude Code, Cursor, or OpenCode installed needs to select one local CLI as the task executor.
- A team with an internal command-line agent wants to integrate it through a custom template using
{workspace}and{prompt_file}. - A domain-expert team wants an Expert-Led workflow where mentor models, users, or human experts participate in real task completion.
- A user needs to audit or reuse prior task output by inspecting, exporting, searching, or pulling Experience Compilations.
How do you install or deploy this agent?
A working Node.js and npm installation, plus the local agent CLI you intend to use, are required. Run:
npx agent-apprenticeship initOr install globally:
npm install -g agent-apprenticeship
apprentice initFor mentor model providers, configure the applicable OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, or OPENROUTER_API_KEY in ~/.agent-apprenticeship/.env.local; the same variables may instead be set for the current shell session. Run apprentice doctor to check setup.
How do you use this agent?
After initialization, run apprentice configure, or use apprentice configure model for the mentor provider, then review settings with apprentice settings. A first task can be started with:
apprentice run "Create a short market map for AI procurement tools."Use apprentice watch <run_id> to monitor it. When it finishes, inspect the emitted Experience Compilation path with apprentice ecosystem inspect <experience_compilation_path>; optionally run apprentice learn install <experience_compilation_path> before a future task. For a session-only loop limit, set AA_MAX_ITERATIONS=3.
What are this agent's strengths and limitations?
- Supports several local agent CLIs in one workflow system and provides a documented custom command-template path.
- Treats completed work as an Experience Compilation that can be checked, fully exported, searched, pulled, and later installed.
- Documents Autonomous, Expert-Led, and Organization Custom Apprenticeship Modes.
- Publishes Seed Dataset v0.2 with 500+ seed tasks, 1,000+ full execution traces, and 505 full experience compilations.
- Requires a local shell, filesystem access, and an installed agent CLI; the documented flow is not a one-click hosted service.
- Mentor configuration depends on API keys for OpenAI, Anthropic, Gemini, or OpenRouter; model pricing and billing are not specified.
- Loop depth must be managed through settings or
AA_MAX_ITERATIONS; no automatic failure-recovery behavior is documented. - The supplied material does not specify collaboration permissions, data redaction, or a review process before contributing to the public ecosystem.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Apprenticeship This agent | 49 · Major gaps | CLIFree + model costs | ★ 1.6k | 18d ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| Agent Toolkit | 71 · Some gaps | CLIFree | ★ 19 | today | V | ChatGPT · Codex · Claude Code |
| AGF Local Coding Session Finder | 77 · Good | CLIFree | ★ 162 | 1mo ago | Rust | Codex · Claude Code |
| GameDesignOS: A Local-First Operating System for AI-Assisted Game Design | 71 · Some gaps | CLIFree | ★ 452 | 2mo ago | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes a local key file, private mode, run artifacts, and public contribution commands; the CLI excerpt also imports sanitization-related modules. These are limited signs of data-flow and sensitive-data handling, so those criteria score 1. The packet does not show concrete least-privilege controls, confirmation before uploads, external-effect safeguards, or rollback procedures, so those scores remain low. The ATIF model names Harbor, pins a commit, and cites its license, earning 2 for source attribution; repository-wide maintainer identity and contribution paths remain unclear.
The test excerpts cover ATIF validation and round trips, artifact integrity, workflows, and rejection of invalid traces, while README commands broadly align with the CLI structure; self-consistency scores 2. Python requirements and dependency ranges are declared, but there is no lockfile or availability assurance in the packet, so dependency availability scores 1. The CLI reports failed or partial tasks and gives reasons, earning 2 for failure messages; other error paths are not sufficiently shown for full marks.
The README lists multiple local agents, model providers, custom commands, autonomous/expert-led/organization modes, and public/private contribution scenarios, earning 2 for audience and scenarios. The CLI offers configuration, diagnostics, run, inspect, export, and learning commands, earning 2 for environment fit. Modes and run settings provide some boundaries, but unsupported scenarios and execution limits are not clearly stated, so capability boundaries score 1. Commands and their intended triggers are clear, earning 2 for trigger precision.
The README organizes the main features and provides installation, quickstart, and command examples, earning 2 for information architecture and install notes. Examples are present but no FAQ is shown, so that criterion scores 1. Package, command, and file naming are broadly consistent, earning 2. No known limitations are documented, scoring 0. The MIT license text is explicit, scoring 3. Package metadata gives version 0.3.0 and the README mentions Seed Dataset v0.2, but no changelog is shown, so versioning scores 1. The repository is named in metadata, but a clear maintainer or update commitment is absent, earning 1 for maintenance responsibility.
The described workflow loops, experience compilation, export, and training reuse are reflected in CLI workflows and test artifacts; the output is inspectable and reusable, earning 2. The packet provides concrete implementation signs for retaining and reusing work experience, earning 2 for marginal value. Economic value and outcome improvement are mostly README claims, with little evidence about costs, resource use, or measured benefits, so cost-benefit scores 1.
README claims about dataset scale and economic value lack corresponding data or methodology details in the supplied files, so claim traceability scores 1. Some architecture and behavior are corroborated across the CLI, package configuration, and tests, earning 2; the test excerpts do not cover all promotional claims. The materials do not consistently distinguish measured facts, estimates, and vision claims, so fact-inference separation scores 1.
- The supplied source sample does not verify the README's dataset-size, economic-value, or performance-improvement claims; unseen files remain unknown.
- The static materials show public contribution and model-key support but do not establish a complete upload manifest, upload-confirmation flow, or full security controls. Review configuration and contribution artifacts before use.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/agent-apprenticeship)<a href="https://followagents.com/en/agents/agent-apprenticeship"><img src="https://followagents.com/badge/agent-apprenticeship.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What is new: - Experience Compiler for iterative extraction of reusable training data - Full training data packages (SFT, reward/verifier training, process supervision, preference learning, and agentic RL) - Runtime training install - Eval, replay, environment, and source evidence map packages - Public Ecosystem and Private Internal Only modes - Iterative workflow loops across local Apprentice Agents - Codex, Cursor, Claude Code, OpenClaw, OpenCode, Hermes Agent, and custom agent support …View on GitHub ↗
FAQ
Which agents can it run directly?
Must experience be contributed publicly?
Public Ecosystem and Private Internal Only modes.What credentials are needed for mentor models?
How is one task's output reused?
apprentice learn install <experience_compilation_path> before future work.What checks are available when setup or output has problems?
apprentice doctor for setup checks and apprentice bundle check <experience_compilation_path> for an Experience Compilation.Thanks — we use this to improve the page.