Second principle: test just the final outcome — treat the agent as a black box. 4:21

Final outcome (black box)

  • Check whether the final result is what you expect: if the user asked to make a booking, did the agent actually complete it? 4:32

Trajectory testing

  • Separately, verify the agent's trajectory — you may know the expected set of steps to complete the task. 4:32
  • Example: the agent must call a list of tools and check APIs in a certain order; test that this sequence was actually followed. 4:44