Learn Harness Engineering
Hands on course for AI coding agents
Description
Same model, different results: some people run Claude Code or Codex for hours and ship a feature, others watch it start mangling files ten minutes in. The difference is usually not the model but the layer around it — the environment it gets, how state is kept, how results are verified, when it stops. Learn Harness Engineering is a project-based course on exactly that layer, harness engineering, teaching you from zero how to make coding agents reliable.
It distills the harness engineering write-ups published by OpenAI and Anthropic, and every lecture comes with a hands-on project, so you finish with a working AGENTS.md, feature list, init script and verification flow rather than a few concepts. Available in 15 languages, with close to 25,000 GitHub stars.
Five subsystems: Instructions, tools, environment, state, and feedback and control — and what breaks when each is missing.
14 lectures, 8 projects: From a first harness to loop engineering (agents rerunning on goals and timers) and graph engineering (specialized, parallel, verified multi-agent work).
Real harness breakdowns: How Claude Code, Codex, Pi and DeepSeek build theirs, such as Claude Code's four-layer memory and sub-agent isolation or Codex's worktree isolation.
Drop-in templates: Goal, loop-state, maker and checker prompt templates you can use right away.
Scaffolding skill: The bundled harness-creator skill generates the skeleton of a production-grade harness.
Free and open: Read it online; MIT licensed.
It distills the harness engineering write-ups published by OpenAI and Anthropic, and every lecture comes with a hands-on project, so you finish with a working AGENTS.md, feature list, init script and verification flow rather than a few concepts. Available in 15 languages, with close to 25,000 GitHub stars.
What it covers
Five subsystems: Instructions, tools, environment, state, and feedback and control — and what breaks when each is missing.
14 lectures, 8 projects: From a first harness to loop engineering (agents rerunning on goals and timers) and graph engineering (specialized, parallel, verified multi-agent work).
Real harness breakdowns: How Claude Code, Codex, Pi and DeepSeek build theirs, such as Claude Code's four-layer memory and sub-agent isolation or Codex's worktree isolation.
Drop-in templates: Goal, loop-state, maker and checker prompt templates you can use right away.
Scaffolding skill: The bundled harness-creator skill generates the skeleton of a production-grade harness.
Free and open: Read it online; MIT licensed.
