How do you catch an agent action that succeeded but was wrong?
A user asks how to catch the sneakiest agent failure: the tool call returns 200, the agent reports success, the logs look clean, but the wrong record was changed. The thread is a practical discussion of verification — checking the end state against what was asked, rather than trusting logs — and keeping a short expected-outcome checklist before running anything risky.