Own the harness, but not all of it

A well-known voice in the agent world made the rounds this week arguing that teams should own their harness: own the context and memory layer, own model optionality, stop renting your whole intelligence stack from one provider. He is right, and the week made his case for him. An open model landed at the frontier’s price, a rival coding tool reportedly grew its usage tenfold, and models keep getting swapped, bundled, and occasionally banned out from under the people building on them. Marry one provider and you inherit its every mood.

The harness, plainly, is the part you control: the tools, the memory, the checks, the glue that lets you point the same workflow at a different model on a Tuesday. Owning it is what turns a model swap from a rewrite into a config change.

Having said that, I would add the caveat the enthusiasm usually skips. Every layer you own is a layer you now maintain. I read about another builder this same week who discovered his lovingly owned setup had swollen into 172 instruction files fighting each other. Ownership is not free. It is a standing bill in attention, and a harness you own but no longer understand is worse than one you rented, because now the mess is yours.

So the real skill is not own everything. It is choosing which layers are worth owning. The memory, the evals, the reversible boundaries around what the agent can touch, those earn their keep because they encode judgment that outlives any one model. The rest you can rent, and should, until it proves it needs a home.

Optionality is the goal, not ownership for its own sake. The point of holding the leash is being able to change the dog. It is not collecting leashes.