People often talk about AI as if the model is the whole product. It is not. The model is the brain. The harness is everything wrapped around that brain: the tools it is allowed to use, the rules it has to follow, the memory it keeps, and the way it shows you what it is doing.
Two products can run on the exact same model and behave completely differently, because their harnesses are different. One might only answer questions. The other might browse the web, edit files, run reports and send messages. Same brain, different harness.
The word comes from the same image as a horse harness: the gear that lets you direct raw power toward useful work.
Why you keep hearing it
As AI moves from chatting to doing (the shift everyone calls "agents"), the harness is where most of the real product work happens. Builders argue about harness design constantly: what the AI should be allowed to do on its own, when it should ask permission, how it should recover from mistakes.
What it means for you
When you are comparing AI tools, remember you are not really comparing models. Most tools use the same handful of models underneath. You are comparing harnesses: what the tool lets the model see, do and decide. That is usually what makes one tool feel smart and another feel useless.