Harness

Everything wrapped around the model (the AI itself — the part that reads and writes) to turn it into an agent (an AI that works towards a goal on its own, step after step). The model by itself can only read text and write text back. It cannot open a file, run anything, or remember what happened a minute ago. The harness is what gives it hands and a memory: it decides what the model gets shown, carries out the actions the model asks for, feeds the results back, and stops to ask your permission wherever you said it should.

The van, not the driver. A good driver with no ladder, no tools and no sat-nav does no work at all. Same driver, properly kitted van, and the job gets done.

Claude.ai — the chat window you already use — is a harness. It happens to be one fitted for conversation: it can search the web and read a file you attach, and that is roughly the extent of it. Put the same model in a harness built for making software and it will edit files on your computer and run commands on its own. Same driver, different van.

Why it matters: same model, different harness, completely different behaviour, even on word-for-word identical instructions. So when an AI product does something strange, the harness is the first suspect and the model is the last — and "which harness is this running in?" is a fair question to put to a supplier. One thing a harness is not is your process. Checklists, review gates and save points are things you build on top of it — see orchestration.

Related: Model · Agent · Orchestration