Source: walkinglabs/learn-harness-engineering README (MIT). Surfaced by Shubham Saboo.
What it is
A project-based course on what you build around a coding agent so it finishes the job reliably: the environment, state management, verification and control. 14 lectures and 8 projects, translated into 15 languages.
The framework
The course uses five harness subsystems: instructions, tools, environment, state and feedback. Everything else is organized around them.
Path through the material
Start with the five subsystems on a real Electron app.
Move to loops where one agent builds and another checks.
Then to agent graphs: parallel work, rollback and a human approval step. Lecture 14 argues why a single loop grows into a graph (prompt, context, loop, graph layers), names failures in-loop checks cannot fix at scale (Goodhart effects, blindness upward, conflicting goals), and asks when a graph is worth its orchestration tax.
A skill scaffolds a harness into your project, and an audit script checks a repo against the five subsystems.
Frontier harness teardowns (August 2026 section)
| Product | Harness choices highlighted |
|---|---|
| Claude Code | four memory layers, five compaction levels, hooks, isolated sub-agents |
| Codex | repository as source of truth, AGENTS.md as a directory page, worktree isolation |
| Pi | minimal kernel, extend it by asking it to build what you need |
| DeepSeek | everything is a plugin, capability seams, an event pipeline |
All four are read through the same five-part lens, so you can compare designs and borrow pieces.
Why it is worth a staff engineer's time
A shared vocabulary for harness reviews, and a checklist via the audit script.
The teardowns are the fastest way to see how production teams trade off memory, compaction and isolation.
The graph lecture is a sober take on when multi-agent structure pays.
Caveats
Teardowns are reverse-engineered from outside; treat details as informed reading, not vendor documentation.
Coding-agent focused; little on non-coding agents or production operation at scale.
Related entries
R023 (building a coding agent from scratch), R087 (Stanford CS329Z), R006 (harness vocabulary), R054, R094.