A generated answer is correct, but an agent repeatedly chooses an unnecessary expensive action before answering. Which improvement should be evaluated?

1. Give the action access to more objects
2. Increase the answer's maximum length
3. Ignore the issue because the final sentence is correct
4. Clarify when the action is needed and review its selection logic using execution traces