The agent did exactly what I asked. That was the bug.

An LLM agent executed user requests correctly but produced incorrect results due to a flawed plan. Automated hooks cannot detect such errors. The author introdu