Harness

A harness is everything wrapped around the model that turns it into a working agent. The model is the brain that makes decisions. Around it, the harness is the rest: the tools it can call, the memory of what's happened, the workspace it can touch, the loop that keeps it running, and the guardrails on what it's allowed to do. There's a shorthand going around for this, agent equals model plus harness.

It's why two apps built on the same model can feel worlds apart. Same brain, very different harness. One has good tools and sensible limits, the other just calls the model and hopes. A lot of what makes an agent feel sharp or clumsy is the harness doing its job, feeding the model the right context and catching it when it slips.

Keep the harness in mind whenever you're comparing AI products, or wondering why one feels so much better than another on the same underlying model. The gap usually lives in the harness rather than the model, so that's where to look. It also explains why "we use Claude" or "we use GPT" tells you almost nothing about how good a product is.

If you build, the harness is most of the job. Picking a model takes an afternoon. Getting its tools and its stopping conditions right is where the real work lives, and it's what turns a slick demo into something you'd trust.