Same Model: Can't Build It for $9, Can Build It for $200
Anthropic ran a controlled experiment: the same Opus 4.5, the same task. Without a harness, the run cost $9 and took 20 minutes, and the output was unusable; with a harness, it cost $200 and took 6 hours, and it produced a game that's actually playable. walkinglabs turned this methodology into a free course: 14 lectures, 8 hands-on projects, 15 languages.
Karpathy once said: the biggest mistake in AI is skipping harness engineering. The quote comes from a post by @RoundtableSpace on X, which also links to a GitHub repo: walkinglabs' free course learn-harness-engineering — 14 lectures, 8 hands-on projects, 15 languages.
Let's set the scene first. You're using Claude Code or Codex on tasks in a real repository. It reads files, edits code, reports progress — everything looks normal. Then you run the tests, and they're red. It says it's done, but the feature doesn't work. You go back and clean up, and it takes longer than writing it yourself would have. A lot of people chalk this up to the model not being strong enough. The course says the opposite: this is a harness problem.

The evidence comes from a controlled experiment by Anthropic. Same model, Opus 4.5; same task: build a 2D retro game editor. Without a harness, it burned $9 in 20 minutes and the output was unusable. With a full harness — planner, generator, and evaluator each doing their part — it burned $200 over 6 hours and produced a game that's actually playable. The model didn't change; what changed was the scaffolding around it. OpenAI reported a similar conclusion with Codex: in a repository with a well-configured harness, the same model went from unreliable to reliable. The course README calls it a qualitative shift — a difference in kind, not degree.
What exactly does "harness" mean
The course breaks it down into five subsystems:

- Instructions: Tells the agent what to do, in what order, and which files to read before starting. Not one giant file, but a map that unfolds as needed.
- State: Records what's done, what's in progress, what's next. Written to disk so the next session can pick up where you left off.
- Verification: If the tests don't pass, it isn't done. The agent can't just talk its way into claiming a win.
- Scope: One feature at a time. No sneaking in three other changes on the side, no deleting unfinished features from the list to hide them.
- Session Lifecycle: Initialize at the start, clean up at the end, and leave a clean startup path for next time.
The course hammers on one point repeatedly: the harness isn't about whether you write good prompts. It's about the environment the model works in.
What's in the course
The whole course revolves around a single product: a personal knowledge base desktop app built with Electron. The 8 projects are the same app at eight stages of evolution — each project starts from the previous project's answer. As the app grows, so does your harness craftsmanship.

Two new pieces of content were added in 2026. Lecture 13 covers loop engineering: moving from manually typing prompts to designing loops — goal loop, timer loop, and maker-checker, with the core idea being to get yourself out of the loop. Lecture 14 covers graph engineering: a loop is just a single-node graph; once a task needs parallelism, shared state, verification, and rollback, it's no longer a loop. August also added Frontier Harness Design Breakdowns, which uses the course's five-subsystem framework to dissect the real implementations at Pi, Claude Code, Codex, and DeepSeek.

What you can use right now
The lowest-effort entry point is four files, dropped into your repo root:
AGENTS.md the agent's operating manual
init.sh install dependencies, run verification, start services
feature_list.json which features exist, which are done
claude-progress.md session progress log
The Resource Library has ready-made templates in all 15 languages, including Simplified Chinese. There's also audit-harness.sh — zero dependencies, no need to install Node.js; just curl it down and it will check an existing repo against the five subsystems to see if it makes the grade.

A few useful replies
Danny B says five lines of CLAUDE.md brought the same task from $0.77 down to $0.46 — same model, same answer.
Someone complained that harness engineering is just babysitting with extra steps. That complaint isn't wrong. But you can't hand off work that needs watching wholesale until you've written the requirements down in files. The act of writing those files is precisely the harness.
Michael Kisilenko points out that the most frequently skipped piece is memory. For small teams, critical context often lives in a WhatsApp group that the agent simply can't see. He built Anyx, which lets the group reach a Grok Bot via email, and the bot answers based on what's been decided in the group. That reply highlights the boundary of the harness: it can govern what's in the repo, but not the consensus reached outside it.
About that $9 and $200
The comparison is easy to read as harness burning more money. Do the full accounting: the $9 version doesn't ship, so the real cost is rework time plus your own hours. The $200 buys something that runs. Which one is more expensive depends on what your hour is worth.
The course has "engineering" in its name, and what it teaches really isn't tricks — it's more like infrastructure. It's for people already using coding agents who keep tripping over the same class of problems. If you're only after prompt tricks, this course isn't for you — the README says so itself.
Repo: https://github.com/walkinglabs/learn-harness-engineering
发布时间: 2026-11-17 02:28