Practical Model Harnesses

Imagine an autonomous information system (AIS). The kind of system Anthropic claims to operate with Claude Code. A system where agents investigate the world (as defined by the system), write prompts for other agents, then evaluate the output of those agents. And improve. In a way, you can think of this system as a Newton's cradle, where the human operator nudges it into motion and it carries on, absent any friction, autonomously, in perpetuity.

Yet, anyone who has deployed such a system knows, without a doubt, that it will degrade over time. Generally, for two reasons:

1. Any amount of ambiguity will resolve to the median of the model's weights. They will narrow information to what is statistically expected absent human intervention and will converge towards the least surprising thing, indefinitely.
2. Models drawn from the same distribution share the same blind spots. A system's confidence in its own errors proliferates over time because there is no external force pushing back. Even if you diversify models as judges, your median only becomes a product of the set of distributions you employ.

This "narrowing" of outputs is actually a feature of LLMs. It is a pressure mechanism, compacting the distributed space of possible answers into what is most likely. When it is only pointed inward, it degrades. But when pointed outward, it becomes a truth-seeking verifier.

When an autonomous information system (AIS) is properly designed, it points this "narrowing" at a goal. But the goal eventually becomes a part of the system and models will converge on a distribution that favors the goal. Left alone, it reinscribes bias.

The AIS is missing one thing: it is not a better model; not a stricter goal; nor a smarter judge. It is you. The Dasein. The "being who is there". You have what the system lacks. What every model needs. A being that has something at stake with the output, for whom the question of what is "good" actually hurts to get wrong.

A model has no stake, and it does not really stand in the face of its own output. We might ask it to. We might prompt it to respond like a human. But beneath this facade, it is only converging on the median of its own distribution.

You have a stake that is not in the distribution of the model's weights. You have to live with the consequences of the model's actions. They may parrot empathy or distress, but in reality they are not like a human is; they do not live with their outputs. They do not persist or live with consequences. Their degradation cannot be arrested from within, as a stake is not a property a model can be given. A model is not Dasein, for it is not being there, it is just there.

Enter: the harness, the apparatus making the human stake legible to the loop. As a model swings across its distribution, our judgment is the standard it is checked against, and the harness is what makes that standard legible. It molds our intent into a framework as it is becoming, not just as it is. It records our stakes and shapes them into a target for the model to aim its narrowing towards. It is capturing a framework as it forms, and that framework is the human made judge.