Harness engineering: every slip makes the harness tighter

When an AI agent makes a mistake, you can respond in two ways. You can fix the result and move on. Or you can ask why it happened and change the agent’s environment, so it does not happen again. The second way is called harness engineering.

The term took off in early 2026. On 11 February, OpenAI described how a small team built a product with coding agents and, by its own account, without writing code by hand. The engineers’ work shifted to designing the environment, the instructions and the checks. On 2 April, Birgitta Böckeler of Thoughtworks published a framework for it on martinfowler.com. Both are about software. The method works for any team.

Guides and sensors

Böckeler sorts every control in a harness into two kinds.

• Guides act before the agent does: instructions, conventions, templates, methods and examples. They aim for a good first attempt.
• Sensors act after the agent does: tests, checks, a second review. They tell the agent what went wrong, so it can correct itself.

You need both. With only sensors, the agent keeps making the same mistake and gets corrected every time. With only guides, you never find out whether your rules work.

Each control can also be computational or inferential. A computational check is exact, fast and cheap: a number adds up or it does not. An inferential check uses a model’s judgement, such as a second AI reading a text against your tone of voice. It catches subtler problems, but it is slower, costs more and is less predictable. So the cheap, exact checks come first.

What guides and sensors look like outside code

In a marketing, sales or operations team, the same two kinds look like this.

• Guides: brand rules, tone of voice, report templates, a written method, a list of what the agent must never do.
• Sensors: every ratio recalculated from the raw numbers, the result held against a quality bar, a second review, a person who approves before anything is sent or published.

A real example: a name that had to be said right

For zaela.com, sixty short films were made with synthetic voices. Zaela is said za-YAY-la, and a speech model does not know that. That was the slip.

• The guide: a phonetic spelling of the name for every language, written down in memory, so every voice starts from the right spelling.
• The sensor: every line that says the name was checked with speech recognition. Where the name was not heard, a verified clip in the same voice replaced it, and the sentence was checked again.

The mistake is now part of the harness: the next film starts with both. Read how the films were made.

Six habits of harness engineering

1. Write it down where the AI can read it. Rules in people’s heads or in old chats do not exist for an agent. In OpenAI’s experiment the code repository was the system of record. For a team, that is shared memory.
2. Break work into small, checkable steps. A step with a clear definition of done is easier to check than one big task.
3. Put cheap, exact checks first. Recalculate, count, compare against a list. Keep a second AI review for what an exact check cannot see.
4. Make every check say how to fix it. “The total does not match the sum of the channels” helps an agent correct itself. “Wrong” does not.
5. Tighten the harness after every slip. Fix it once, in the instructions, instead of again and again in the chat. Google’s developer team advises starting with one recurring failure.
6. Bound what the agent can touch, and keep the logs. Access per person and approval for risky actions belong to the harness. Logs show what happened, and they are what you learn from.

The limits

Harness engineering is not a guarantee. Böckeler is clear about what sensors miss: a problem that was misdiagnosed, a solution that is more complex than it needs to be, an instruction the agent misunderstood. And OpenAI’s experiment is one self-reported case, on a new product without old code. Keep a person at the gate where taste and responsibility matter.

Harness engineering in Zaela

Zaela is a company AI harness, built so a team can do this work without writing code. Guides live in memory, in skills and in the instructions of each agent. Sensors are part of the methods: exact calculation, a quality bar, a list of what an agent never does, approval before anything goes live. Every action is logged. And everything the AI uses, you can see, adjust and add to in the web app, so a lesson goes into the harness the same day.

The rule of thumb is the same as for agents. Ask for the same thing twice? Turn it into an agent. Correct the same thing twice? Turn it into a rule. For the full background, read the guide to harness engineering on i-scoop.eu, also published by Edge.