Lesson 01 · Foundations · 13 min

Humans on the loop

Before designing loops it is worth deciding where you intend to stand relative to them. Kief Morris describes three positions, and the middle one is the trap.

Outside the loopYou state a goal and let the agent do everything. Fast when it works. When it does not, the agent spirals and the cycle takes longer than doing it yourself.

In the loopYou inspect and correct every step. Safe, and a bottleneck: agents produce work faster than a person can read it, so the gain evaporates.

On the loopYou build the harness — the specifications, the checks, the guidance — and let it govern the work. When you are unhappy, you change the harness rather than the artefact.

In the loop, you fix the artefact. On the loop, you change the harness that produced it — and every future output moves with it.

That last distinction is the whole discipline in a sentence. In-the-loop, a bad output means you fix the output. On-the-loop, a bad output means the harness let it through, so you change the harness and every future output moves with it.

Morris splits the work into two loops that belong to different parties. The why loop is human: what are we building, for whom, and what would count as good. The how loop is the agent's: the implementation, at whatever depth. Confusion between them is what produces both micromanagement and abdication.

Being on the loop is not less work. It is different work, and it is harder to fake — writing a check that captures what you actually mean by "good" is more demanding than noticing that one particular reply was wrong.

In the lab, the rubric is your position on the loop. It is not a prompt to the writer; it is the standard the checker enforces on your behalf, every time, whether or not you are watching.

Checkpoint

Find a reply the checker passed that you would have rejected, and say what the rubric was missing.