What is harness engineering? A guide for small business

Harness engineering is setting up an AI model so it does one job your way, every time: written instructions, a short list of tools it may use, checks before anything leaves, and a rule added after every mistake. The model does the thinking. The harness decides what it is told, what it can touch and what gets caught.

An AI model on its own answers whatever it is asked, as well as it can, once. A business needs more than that: the same job done the same way every day, with nothing sent to a customer that should not be. The harness is how you get there.

The model is the engine

ChatGPT, Claude and Gemini are models. They are very capable, they improve every few months, and you control almost nothing about how they are built. Choosing one matters less than most people think, and we cover how to choose in which AI model a small business should use.

The harness is everything around it

Written instructions for the job. A short list of tools the AI may use, and nothing else. Checks that run before anything leaves the building. And a habit: when the AI gets something wrong, you change the harness so it can't happen again. None of these parts is clever on its own. Together they turn a capable model into a dependable member of staff.

Where the word comes from

Harness is an engineer's word for the frame that runs and tests something. The AI companies now use it for exactly this. Anthropic, the company behind Claude, calls its own agent kit a harness in its guide to effective harnesses for long-running agents, and the tricks in that guide are plain ones: a progress file the AI reads at the start of each session, one feature at a time, and testing tools so it can check its own work.

A worked example: a plumber's enquiry replies

Take a two-van plumbing firm in Leeds that gets 15 to 20 enquiries a week. Without a harness, the owner pastes an email into ChatGPT at 9pm and hopes. With one, the AI reads each new enquiry against a one-page brief, can only write drafts and never send them, has the real price list and call-out policy to hand, and flags anything about a gas smell or a complaint for the owner. Same model. A very different Tuesday.

This is our whole approach in one line: model, then harness, then reliable output. The magic happens in the middle, because that is the part you can change today and the part that stays when the model changes.

Same model, different results

Give two businesses the same model. One types a question into a chat window when it remembers to. The other has the AI read the morning's emails, sort them by a written rule, draft replies in the house style and hold anything about money for a person. Same engine, a different business.

Most failures are set-up, not intelligence

When AI gets something wrong at work, it is usually because it was never told the rule, was given a tool it should not have had, or had nothing checking its work. Those are harness problems, and they are fixable today. Making things up is the big one, and most of the cure is in the harness too, as why AI makes things up explains.

Every mistake becomes a rule

A harness gets better every week if you treat each mistake as a missing rule. Write the rule down, add the check, and that mistake is gone for good. Over a few months, the AI stops feeling clever and starts feeling dependable. The habit has its own guide: every AI mistake should become a rule.

Models change, the harness stays

New models arrive every few months. If your set-up lives in your head and a chat history, every switch starts from zero. If it lives in a written brief, a tool list and a set of checks, you move them across and test. The harness is the part of your AI set-up you actually own.

The most useful thing to understand about a harness is the difference between asking and enforcing. Instructions ask. Tool limits and checks enforce. A good harness uses both, and puts anything that would be expensive behind the second kind.

Why a written rule is not enough

Even the AI companies say so. Anthropic's documentation for Claude Code says its written project instructions are treated as context, not enforced configuration, and that to block an action regardless of what the AI decides you need a hook: a small piece of code that always runs, described in its guide to hooks. The same holds in a small business. "Never send without approval" is a rule. Draft-only access is a lock.

Design the system, not just the prompt

The UK's National Cyber Security Centre makes the same point about AI generally: no model works in isolation, so design the whole system with its risks in mind, for example with a rules-based layer that stops damaging actions even when the model is talked into them. That is in its post on thinking about the security of AI systems.

Keep people in control at the right points

The UK government's own AI Playbook puts it as a principle: meaningful human control at the right stages. For an owner-run business, the right stages are anything that sends, pays, deletes or promises. Which AI jobs still need a person to sign off goes through them.

For one well-defined job, expect a few hours to set up and a few minutes a week to keep. Most of the setting-up is writing down what you already know. Most of the keeping is adding a line when something goes wrong.

Setting it up

The brief takes an hour or two if you write it from real examples. Connecting the tools with the right limits takes an afternoon, longer if your systems are old. The trial runs alongside your normal work for a fortnight, so it costs attention more than time.

Keeping it running

Ten minutes a week reading the mistake log and adding rules, for the first couple of months. After that it settles, and you mainly come back to it when a price, a policy or the model changes.

Where the time comes back

Example figures, not a promise: if each enquiry reply takes you 8 minutes by hand and 2 minutes to check a good draft, 18 enquiries a week saves you about an hour and three quarters. That only holds if the drafts are good, which is the whole job of the harness.

Four questions tell you. If you answer no to any of them, that is the next piece to build.

Is the brief written down where the AI reads it every time?

Not in your head, not in one long chat. A saved project, custom assistant or settings page, so every job starts from the same page.

Could you list everything it can touch?

If you don't know whether it can send, delete or pay, assume it can, and check its connections today.

Does anything stand between the AI and a customer?

A person, a rule-based check or both. If nothing does, you are trusting the model alone, and that is the set-up a harness exists to replace.

Do you keep a log of its mistakes?

One line each: what happened, why, and the rule it became. No log means the same mistakes will come back.

A harness makes good work repeatable. It can't make a broken process good, and it can't invent facts you never gave it.

A process that doesn't work by hand

If the job is muddled when a person does it, the AI will be muddled faster. Fix the process on paper first, then hand it over.

Information you never gave it

The AI only knows what it can see. If your prices, policies or opening hours aren't written down somewhere it can read, it will guess, and a harness should stop it guessing. Our Knowledge Index maps where every document lives, so the AI reads the current one.

Jobs the model can't do yet

Some jobs are beyond today's models however good the set-up: anything that needs a site visit, a judgement call on a person, or a fact nobody has recorded. What AI can't do for your business lists the honest limits.