Where the human goes

The person never left the factory. They moved — and the job at the new position is a different job.

The anxious version of this story has people being removed from the line. The observable version is that they moved to the end of it.

Agents own the inner loop: investigate, implement, test, report. Engineers own the outer loop: decide whether this is the right change at all, verify that the diagnosis is sound, approve, and carry the consequences of being wrong. Those are not the same skill, and the second one does not get easier as the first gets faster — it gets more consequential, because there is more of it arriving per hour.

the agent's work the engineer's work, and what crosses between

Fig. 7.1The person did not leave the factory. They moved to the end of the line, where the evidence arrives.

The boundary is evidence

What crosses between the loops is evidence, not a claim: the diff, the tests that ran, the logs, and a short explanation that connects them to each other. An agent reporting “fixed the race condition” has given you nothing to act on. An agent reporting the diff, the failing test that now passes, the log line that identified the interleaving, and two sentences tying them together has given you something you can approve or reject in ninety seconds.

Designing that report is real work — harness work rather than model work. The teams getting the most out of factories are largely the teams that made the handoff artefact good.

The same shape shows up outside code, which is a decent sign it is structural rather than faddish: a few specialised agents in sequence — researcher, writer, critic, publisher — with the human approval sitting at the point where something becomes irreversible.

3–7 agents per enterprise workflow, with human review at side-effect boundaries as of 2026 Reported

Broad reads, narrow writes, a person at the point of no return. That is the pattern, whether the artefact is a pull request or a published page.

The job at the new position

Three things change about the work, and none of them are “less engineering”.

You are designing the line, not working it. The leverage moved from writing the change to deciding which changes are possible, cheap and safe to make. Types and signatures so the compiler catches what the agent gets wrong. Test seams. Short call stacks and legible layout, because an agent reads code the way a new hire does. Component boundaries that keep a mistake contained. Dependency injection. None of this is new — it is the same architecture advice as always — but it is doing a second job now: it is a cheap, hard-to-fake safety net against mistakes a machine will make at volume.

You are accountable for output you did not write. This is the uncomfortable part, and no amount of tooling relieves it. Approving is authorship. The legal and professional weight of a change sits with whoever merged it, and the fact that a machine wrote it does not distribute that weight anywhere else. Reported

Your intuition about your own speed is not evidence. The perception gap from the METR trial Contested applies to you personally, and it applies most strongly when you are enjoying yourself. Measure throughput and defect escape rates; do not trust the feeling of velocity.

From the floor

The last human gate in this factory is a feel-check: after both judges pass, a person reads the site and says whether it works. It is defined as the residue — whatever the judges cannot see.

I would have bet that gate was ceremonial by now. It is not. On the most recent run, both judges passed, and the feedback round that followed still changed things:

1 human feedback round that still caught something after both judges had passed as of 2026-07-23 First-hand

That one round is the reason the gate is still there. It is also the measurement that matters for anyone thinking about darkening. Deleting a gate does not make its residue zero. It makes it invisible. The only way to retire a human gate is to keep it for a while and count what it still catches once the machines are satisfied — and to retire it when that count has been zero for long enough that the number means something, on that kind of work, with evidence rather than enthusiasm.

Done properly, darkness is the asymptote of an outer loop, not a switch you flip.

Evidence: website-factory/specs/retro/2026-07-23.md, § “Verdict statistics”

Where this leaves you

Robots are fine in the dark. People need to see what they are doing — and the engineering question of 2026 is not how to remove them from the building, but which few places they have to be standing for the whole thing to stay trustworthy.

That is a design problem, it is yours, and no model release is going to solve it for you.