The harness has an expiration date. Ours mostly doesn’t.

There is a line going around that we keep nodding at: as the models get stronger, the harness around them matters less. The scaffolding you built to make a weaker model reliable — the prompt chains, the routing layers, the retrieval bolted on to paper over a short context, the loops that catch the model’s mistakes — is a tax you pay for a gap in capability, and each new generation closes some of the gap. People have started treating this as the bitter lesson arriving for agents: every piece of harness logic should carry an expiration date, and when the next model can do the thing without your scaffolding, you delete the scaffolding. Teams have been deleting in public — tools removed, retrieval systems torn out, whole routing layers replaced with a plain handoff.

We think that is exactly right, and also that it is being used to argue something false. There is a competing claim, just as loud, that the harness is now the lever — that in 2026 the marginal reliability of an agent comes more from the harness than from swapping in a slightly better model. Both camps are describing real things. They are just not describing the same harness.

Because “harness” is two different objects wearing one word. One kind exists only to compensate for what the model cannot yet do on its own. The other exists to hold what the model will never do for itself, no matter how smart it gets. The first has an expiration date. The second compounds.

The compensating harness is the one the bitter lesson is right about. Prompt scaffolding, error-correction loops, elaborate routing, retrieval standing in for a weak context window — every one of these is a bet that the model is worse than it will be next quarter. Treat that code as a consumable. We write it freely and we delete it without ceremony; a skill that existed only to walk a weaker model through a task gets thrown out the week a stronger model can do the task cold. Keeping it would be sentimentality, and sentimental scaffolding is how a system rots.

The durable harness is a different animal, and it is most of what we actually built. It is not there to make the model smarter. It is there to make the company real. A file-based brain that an agent maintains so that what we learn on Tuesday is still known in March — no model ships with your institution’s memory. A scheduled heartbeat that keeps un-forgetful attention on the things that decay when no one is watching — no model volunteers a cadence. Gates on the two verbs that can hurt you, spending money and speaking outside the company — no model should be trusted to grant itself those, and a smarter one only raises the stakes. And, from the last essay, an evaluator kept structurally isolated from the thing it scores — a check the checked cannot touch. None of these are crutches for a weak model. They are the institutional layer, and a better model makes them more valuable, not less, because it does more per day and so needs a stronger memory, a firmer gate, a cleaner ledger.

The question was never how much harness. It is which harness — the kind a smarter model deletes, or the kind a smarter model needs more of.

This is why our operating system does not look like a pile of prompt engineering, even though we are two engineers who could have built exactly that. The parts we invest in are the parts with no expiration date: memory, cadence, permission, accountability. The parts we hold loosely are the ones a model release can obsolete overnight. When a new model lands, our first move is subtraction — we go looking for the scaffolding it just made pointless and we remove it — and our second move is to notice that the durable layer now carries more weight, because the agent wearing it just got more capable of doing damage as well as work.

So we nod at the line, with an edit. As the models get stronger, the harness that apologized for them matters less, and you should be deleting it on a schedule. The harness that remembers, gates, paces, and holds to account matters more. A company that only heard the first half will keep tearing out scaffolding until it has torn out its own institution. We are trying to build the half that survives the next model — and the one after that.