When teams talk about their agents, they talk about the model. Which checkpoint, which provider, which prompt. But sit with an agent product for a week and you realize the model is the smallest variable in the system. The harness is doing most of the work.
By harness I mean everything around the model: the tool layer, the sandboxing of what those tools can do, the permission model that gates risky actions, the memory and context window management, and the observability that lets you see what actually happened on a given turn.
These are software engineering problems, not ML problems. They look unglamorous on a roadmap. They are also where every honest AI product team spends the bulk of its time. A weaker model in a well-designed harness routinely beats a stronger model dropped into a thin wrapper.
The lesson I keep relearning: when an agent feels magical, the harness is invisible to the user but obsessed over by the team. When it feels broken, the harness is what shipped first and was never finished.