# Learn Harness Engineering

A project-based course on building the environment, state management, verification and control mechanisms that make AI coding agent harnesses work reliably.

Learn Harness Engineering is a project-based course, translated into fifteen languages, on building the environment, state management, verification and control mechanisms behind reliable AI coding agents. It teaches a five-subsystem framework covering instructions, tools, environment, state and feedback, and applies that framework in a set of breakdowns that reverse-engineer how existing products such as Pi, Claude Code, Codex and DeepSeek build their own harnesses, looking at details such as Claude Code's four-layer memory, five-level compaction and sub-agent isolation, or Codex's use of the repository as source of truth and worktree isolation.

A later part of the course covers graph engineering: why a single agent loop tends to grow into a graph as a project scales, the four stacked layers from prompt to context to loop to graph, the parts of a graph such as nodes, edges, shared state and routing, and why simple in-loop checkpoints cannot fix certain structural failures at scale. An accompanying project has the reader draw their own maker-checker workflow as an explicit graph, then extend it with a parallel fan-out and fan-in step and a conditional rollback path.

Website: https://walkinglabs.github.io/learn-harness-engineering/