An agent returns a wrong answer. A trace shows an incorrect tool argument even though retrieval and model serving both returned HTTP 200. What should the team investigate first?

1. Whether every tool call can be removed without changing the task.
2. Only whether the endpoint received more than one request that day.
3. Only the storage size of the document corpus.
4. The argument construction and validation around that tool call.