Evaluating agents beyond task success: Anticipate and Verify - Behmaneshrad