AI and LLMs · Senior · Question 33 of 75
How should agents be evaluated beyond the final answer?
- 1Only check the final answer, since steps don't matter
- 2Check the trajectory: tools, arguments, no harmful steps
- 3Count how many words the agent writes in total
Show the answer
Correct answer: Check the trajectory: tools, arguments, no harmful steps
Explanation
An agent may reach the right answer by luck or via unsafe actions. Trajectory and end-state checks catch both.