A human in the loop is a gating mechanism — it caps the loop at human waking hours

Most companies experienced ChatGPT as a question-answer bot and then graduated to longer-running agents that go away, get stuck, and come back. Blomfield’s observation is that this shape makes the human “the gating mechanism”: if the agent stops halfway through at three in the morning waiting on input, it cannot go anywhere. The productivity framing — engineers 20% more productive, copilots for lawyers — hides the structural fact that every one of those gains still runs at human speed, because a person sits in the path of every decision.

The prescription follows directly: if you can close the loop without a human, the product improves itself while you sleep. This is why The quality gate should be a second adversarial model, not a human is load-bearing rather than a cost optimisation — the human gate is the thing that has to go for Without a learning mechanism you have an AI pipeline, not an AI loop to matter. It sharpens The gains come from redesigning work around AI, not bolting AI onto human workflows with a concrete diagnostic: ask where in your system a person is holding the baton overnight. Note the deliberate tension with Humans should supervise agent loops from a leveraged point, not sit inside every one — both agree the human should not sit inside the loop, but disagree on how much supervision the leveraged point still owes.