The Invisible Layer That Turns AI Into A Superpower

Everyone obsesses over smarter AI models. The people shipping AI that actually works obsess over the harness: the layer around the model that makes it reliable. This lesson explains what a harness is, the 5 parts it breaks into, and the exact templates to copy, pulled from the free WalkingLabs Harness Engineering course.

What You'll Learn

• Why capable agents still fail, and why it's almost never the model
• The 5 parts of a harness, in plain English
• The exact templates to copy: CLAUDE.md, progress file, evaluator
• The free course plus a skill that builds the harness for you

The Big Idea

An agent is the model plus the harness. The model is the brain: it interprets context and makes decisions. The harness is everything around it: the tools it can use, the memory it keeps, the loop it runs in, and the checks that catch it before it goes wrong. Everyone is racing to build smarter brains. The people actually shipping reliable AI are quietly engineering the harness. The framing comes from @santtiagom_; the playbook below is the free WalkingLabs course.

The line that says it all: Anthropic ran the same model on the same task twice. Bare: 20 minutes, 9 dollars, broken. With a full harness: a fully playable app. They never changed the model. As the course puts it, when things fail, don't swap the model first, check the harness.

Why Your Agent Keeps Failing

It's almost never that the model isn't smart enough. A top model still fails on real, messy work because it's missing the support around it. The course's core principle: model capability and execution reliability are two different things. Fix the harness before you pay for a smarter brain.

Typical failure modes:

• It declares victory too early (says done when it isn't)
• It loses the thread on long tasks across sessions
• It overreaches: tries to fix five things and finishes none
• It has no way to actually verify its own work

What A Harness Actually Is

Rule of thumb from the course: if it's not the model's weights, it's harness. It breaks into 5 parts. Feedback is the highest-value one, and it's the part almost everyone skips.

1. Instructions: clear rules and boundaries the agent actually follows (your CLAUDE.md / AGENTS.md)
2. Tools: what it can do: read files, run code, call APIs, hit the database
3. Environment: a runnable setup so it can actually execute and test, not just talk
4. State / memory: progress files and decision logs so it remembers across sessions
5. Feedback: tests and an independent checker that catch it before it ships garbage. Highest ROI.

The Templates To Steal

The course ships these as blank, copy-ready files. Drop them into your repo and your agent instantly has a harness.

• CLAUDE.md / AGENTS.md: a map, not a manual. Roughly 100 lines: what the project is, how to run it, how to verify it, the non-negotiable rules
• progress.md: done / in progress / blocked / next, plus current commit and test status, so it never loses the thread
• session-handoff.md: a clock-out note for the next session: what's verified, what changed, what's broken, the next best step
• evaluator-rubric.md: an independent checklist that grades the work Accept / Revise / Block, so the agent can't grade itself too kindly

The one rule that fixes the most: Make the agent prove it's done with a real command, not a vibe. "The code looks fine" doesn't count. "It runs and the test passes" counts. Separate the worker from the checker.

The Shortcut: A Skill That Builds It

If you don't want to assemble it by hand, the course ships a skill that creates and improves the whole harness for you. One command:

npx skills add walkinglabs/learn-harness-engineering

Pick the harness-creator skill when the installer asks which skill to add.

Receipts, the big players agree: Andrej Karpathy created AGENTS.md, the instruction file the agent world standardized on. Google just shipped a harness called Antigravity at I/O. OpenAI literally coined the term harness engineering. This is where the field is going.

How I'd Start This Week

1. Add a CLAUDE.md. Map your project in roughly 100 lines: what it is, how to run it, how to verify it, the rules it must never break.
2. Add a progress file. Make the agent update it every session so it stops forgetting across long tasks.
3. Make it prove done. Require a real run or passing test before it's allowed to say a feature works.
4. Add an evaluator. A separate checklist or second agent that grades the work, so the builder isn't grading itself.

The free course: the full WalkingLabs Harness Engineering course is 12 lectures, 6 hands-on projects, and copy-ready templates, free and open source at https://walkinglabs.github.io/learn-harness-engineering/en

Go Deeper

Printable PDF version of this guide: https://drive.google.com/file/d/1_ZA9rs4bB3GuwFcUP-ZnWkBuW8ZzpPna/view?usp=drivesdk

Want the whole system installed live, with Cooper on the call? Build Your Own AI, the two week sprint runs the first two weeks of August.

Or get a new working tool every Monday inside The AI Operator community on Whop.