An AI agent I built gave a fluent, confident, totally wrong answer to a customer - because it trusted stale data without knowing it was stale. This post breaks down what actually went wrong (it wasn't the model), why "confident and wrong" is a uniquely dangerous failure mode for AI agents, and the simple fix that caught it.