Learn Harness Engineering

Learn Harness Engineering is a course from the walkinglabs organization on engineering the environment around AI coding agents such as Codex and Claude Code. It combines 14 lectures, 8 hands-on projects and a copy-ready resource library, published as a VitePress documentation site and an MIT-licensed GitHub repository in 15 languages.

What It Is

The course teaches harness engineering: designing the working system a model operates inside rather than just writing better prompts. It defines a harness through subsystems including instructions, state, verification, scope and session lifecycle. The course draws on published work from OpenAI, Anthropic and others.

How the Course Is Structured

Lectures explain why capable agents still fail, why the repository must be the system of record, why one giant instruction file fails, why long tasks lose continuity, and how end-to-end testing and observability change results. Later lectures cover loop engineering and graph engineering. Projects apply these ideas to a single capstone app, an Electron-based personal knowledge base, comparing prompt-only runs with rules-first harnesses.

Templates, Skills and Tools

The resource library offers starter files such as AGENTS.md, feature_list.json, init.sh and claude-progress.md that can be dropped into a repository. The repo also includes a harness-creator skill for scaffolding a harness and a shell script, audit-harness.sh, that audits a repository against the harness subsystems without needing Node.js.

Frontier Harness Design Breakdowns

A newer section applies the course framework to how Pi, Claude Code, Codex and DeepSeek design their harnesses.

Requirements

Completing the projects requires a coding agent, such as Claude Code, Codex or another IDE or CLI agent, that can edit files and run commands in a local repository.