How to detect agent misbehavior in production at scale

Behavior deviation detection, reasoning artifact capture, and adversarial prompt detection run continuously so human reviewers act on signals, not haystacks.