harness

A harness is the layer between you and the model. It decides what context gets sent along with your request, what tools the model may call, how many attempts it gets, what happens when the conversation grows too long, and increasingly which model handles which part of a job. The model is the engine, the harness is the rest of the car.

This matters more than it sounds. The same model can score five times higher or lower on a test depending only on the harness around it, because things like keeping or discarding its chain of thought between steps change what it can actually do. Companies are also using harnesses to save money, routing routine work to a small cheap model and calling in an expensive one only for the hard parts. When people talk about orchestrators or agent frameworks, this is what they mean.