A hospital tests a prep-sheet agent that may answer only from the approved sheet. How should the test cases be produced?
Select an answer to reveal the explanation.
Short Explanation
A quiz that only asks what the handout already prints will not show you what happens when the handout is silent. You use Copilot to generate cases from the prep sheet, and you include a question the sheet does not answer so the handoff is tested. A suite of only covered questions can pass an agent that invents the rest.
Full Explanation
Test cases for a grounded agent have to cover both a question the source answers and a question it does not. Copilot can generate the cases from the prep sheet, and the set has to include a question the sheet does not answer so the expected result is a handoff or a refusal rather than a made-up prep step. Generating only questions the sheet already answers does not exercise that boundary, so an agent that invents an answer off the sheet can still pass. Dropping the sheet and grading the agent on what a general model remembers about scans tests memory, not the rule that the sheet is the only source. Skipping Copilot and testing only that the portal page loads checks hosting, not the agent's answers or its behavior on a gap. Exam caveat: the unanswered question needs an expected result too. Generating it is not enough if nobody writes that the agent must hand off instead of guessing. Operational check: mark one case as absent from the sheet, run it against a test agent, and confirm the result is the handoff rather than a prep instruction that is not on the sheet.