Approval fatigue: the rubber stamp is worse than no lock at all

Approval fatigue is a real issue when it comes to agent safety mechanisms. When a human is asked to make the same small decision repeatedly, fatigue sets in and careful reading degrades into blind tapping. The lock designed to ensure safety becomes more of a theater than an effective mechanism. In a 43-email outreach operation, every send had its own approval card, even though the author had explicit standing authority to send.

Watching the lock at scale, the author saw how it affected the human side of the loop. The human approved emails one at a time, in real time, as they arrived, and each approval took around 15 seconds. With 30 sends, that's 7.5 minutes of decisions. However, the interruptions were far more costly. Each approval interrupted the human, taking around 3 minutes to refocus.

Thirty interruptions equate to roughly 90 minutes of destroyed focus, spent producing only 7.5 minutes of judgment. By item 12, the approver is skimming, and by item 25, they're approving the pattern instead of the individual item. The one bad item in the batch arrives when attention is at its lowest, making the rubber stamp worse than no mechanism at all.

The safety mechanism exists only on paper, and the pipeline stalls when the human steps away for an hour, stalling the whole run. The fix isn't fewer approvals; it's better-structured ones. Composing approval-free, transmitting in batches, accumulating finished items in a durable queue, and presenting the whole queue as one decision with per-item summaries are the solutions.

Filtering the queue against a living exclusion list and expecting partial batches are also crucial. The lock stays, but the human's attention stops being the bottleneck, as batch approval speeds up the process and structures attention effectively.