On the Context Harness.

There is a moment, when you have built with these models for long enough, that the question quietly inverts. You stop asking what the model can do and start asking what it has been given. The locus of attention shifts outward, from the engine to the harness — the small architecture of tools, memory, instructions, and retrieval sitting between the model and the world. What once felt like periphery turns out, on closer reading, to have been the work all along.

By harness I mean something more specific than scaffolding, though the field is still settling its terms. It is the loop that calls the model, parses its output, dispatches its tool calls, and decides when to stop. It is the rules for what gets remembered between sessions and what is allowed to be forgotten. It is the system prompt, the retrieval index, the schemas that constrain what the model is allowed to say back. It is, roughly, everything in the system except the model itself — which is to say, almost everything.

Philosophers have a name for the inversion. Heidegger called the everyday objects we use ready-to-hand: present not as objects of thought but as extensions of the act being performed. The hammer disappears into the hammering; we notice it only when it breaks. A model, likewise, fades into the workflow that holds it. We have spent two years staring at the hammer. The harness is what does the building.

It would be tempting to call this harness a kind of memory, but the analogy misleads more than it helps. Memory, in us, is reconstructive — biased toward meaning, shaped by what we cared enough to keep, distorted in ways that often turn out to be useful. A harness is curatorial. Someone has decided what to preserve, in what form, with what provenance, surfaced under what conditions.

It is closer, in this sense, to an archive than to a mind. And an archive, as Derrida liked to insist, is never neutral. It is bound up with the question of who keeps it and on whose authority. It encodes what its keepers believed was worth keeping. The phrase context engineering, now widely circulated, papers over an older and harder question: who decides what gets to count as context?

This is why the harness is becoming the product. The model itself is drifting toward commodity; the harness is where a company's judgment now lives. Which evidence is admissible. Which interruptions are tolerable. Which actions require a second pair of eyes. Which silences are worth breaking. Software has always encoded judgment, but for the first time we are letting that judgment act quietly, on a user's behalf, without first being translated into a button to be pressed.

“Most current AI products work in a void and call the void general intelligence.”

There is a humbler way to put it. The harness is where the model meets the situation, and a situation is not a prompt. It is a posture, a set of constraints, a history of what was tried last week, a sense of who is in the room and what they came in carrying. To be, Heidegger thought, is to be situated. To work without situation is to work in a void; most current AI products work in a void and call the void general intelligence.

The literature has begun to acknowledge this in its own dialect. Even with context windows in the millions, practitioners report what they have started to call context rot: as the window fills, instruction-following decays and the model's grip on its task loosens. Bigger contexts, it turns out, are not better contexts. The size of the room is not the question. The question is what furniture has been arranged in it, and where the model has been asked to stand.

There is an older argument worth recalling here. In 1998, Andy Clark and David Chalmers proposed that the mind extends past the skull — that a notebook, in the hands of someone who relies on it, becomes part of how that person thinks. The harness is the analogous structure for a model. But the notebook in their parable belongs to the person who carries it; the harness, in most current products, is the curation of a third party. We are building extended cognition on borrowed notebooks.

I do not think this is, in itself, a problem. Most of cognition is borrowed; libraries and institutions and operating systems are also kinds of harness, and we have spent centuries making them tolerable. But the question of whose interests the curation serves is unresolved, and it will not be resolved by better retrieval. It will be resolved, if at all, by the slow accumulation of norms about what a system is allowed to remember, on whose behalf, and at whose request.

The work, then, is not to wait for the next checkpoint but to thicken the harness — to give the model a place to stand. Give me a place to stand, Archimedes is supposed to have said, and a lever long enough, and I will move the world. We have the lever. The place to stand is what we are still, slowly, learning how to build.