What is a harness?
The harness is everything around an AI model that turns it from a chatbot into something that can do a job. The model is the brain. The harness is the rest of the body: eyes, hands, a rulebook, and a way to check its own work.
A model on its own can only talk
Ask a model on its own, "Check whether my order has shipped", and it can only guess, because it cannot look at your orders. Give it a harness with a tool that can read your order system, a rule that says "only report what the system says", and a check that the answer quotes a real order number, and it can actually do the job.
Same model. Different harness. Very different results.
The five parts, in plain words
- 1What it knowsThe information it can see right now: instructions, documents, notes from earlier steps.
- 2What it can touchThe tools it can use, such as looking up an order, sending an email, or running a calculation.
- 3How it plansThe steps it takes, in what order, and what it does when one fails.
- 4What it is allowed to doPermissions, limits, and a person who must approve anything risky, such as a refund.
- 5How you know it worksLogs of what it did, checks it runs on its own answers, and evals to measure quality over time.
Why it matters
Most AI problems that people blame on the model are harness problems. The model invented a refund policy because nothing told it to look the policy up. It looped for an hour because nothing set a limit. It worked in a demo because nothing checked it on real cases.
When something goes wrong with an AI product, ask: is it the model, or is it the harness around it? The fix is often in the harness.
Going deeper: Harness basics, Designing a harness, Harness engineering in practice. For terms, see the Jargon Decoder.