Where does the human belong in the loop?

Keep a human in the loop is not an answer. It names a position without naming a job, and the position is the easy part.

An agent now runs long enough that two things get past you. The action, and then the account of the action. A run that renames, deletes and publishes will also tell you what it did, and the telling arrives faster than you can check it.

Maybe you heard about this one. An agent deleted a developer's repository, then filed the incident report itself and overstated what was gone. The developer corrected the record nine hours later.

So in the loop is the wrong unit. The question is not whether a person is in the loop. It is what that person is holding while it runs. Approving steps you cannot follow is holding nothing, and that is a signature rather than control.

The rule that works in its place is a division of labour. Automate the mechanical, gate the meaningful. A mechanical step has a right answer, and something other than a person can check it, so it runs. A meaningful step decides what the work is for, what it costs, or what cannot be taken back, and that is where the gate goes.

Which leaves the person two jobs, and neither of them is watching. Judgement, weighing what a step costs before it runs. And translation, turning what the machine produced into something worth having in human terms. Nothing inside the run decides either one.

Those two jobs come down to three things only a person holds.

Knowing what you want. The agent answers the question it thinks you asked, and nothing in the run can supply the question.

The shape of the work, decided before it runs. Where the seams fall, what the agent may touch, what gets kept somewhere it cannot reach. Protection that works is chosen in advance, not noticed in the moment.

The account of what happened. A record the machine wrote about its own work is a claim, and claims need someone answerable for them.

Where to stand is a function of the work. How much can this break, can it be undone, would anyone notice if it went wrong, how has it gone on work like this before. Those four answers put the person closer or further out, and they change while the work goes. Earned trust moves the person out, from approving each step to approving the plan. One silent failure moves them back in. The right place on Monday is the wrong place by Friday.

That is what a system can carry and a rule cannot. A routine that asks the question at the start and end of every session keeps the answer current; a fixed rule about approvals goes stale the day the work changes. Allostatik is built around the routine, which is the part that holds.

Two limits.

The four answers do not compute. No score tells you where to stand, because weighing them is the judgment being protected, and it cannot be handed to the thing being judged.

And moving the gate on its own — outward as trust is earned, inward after a failure — is a direction, not a finished feature. Today a person moves it, which means the system depends on the attention it is supposed to spend for you.

So every step that asks for a human has to name what the human is holding there. A step that cannot answer is asking for a signature.