{"id":1844,"date":"2025-07-30T06:20:36","date_gmt":"2025-07-30T10:20:36","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/powerplatform\/?p=1844"},"modified":"2025-07-30T06:20:36","modified_gmt":"2025-07-30T10:20:36","slug":"plan-validation-cat-kit","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/powerplatform\/plan-validation-cat-kit\/","title":{"rendered":"Introducing Plan Validation in Copilot Studio Kit"},"content":{"rendered":"

Ever wonder why your agent’s answers can be right but for the wrong reasons? Let’s dive into why that matters, with a brief foray into epistemology (bear with me!)<\/p>\n

Knowledge: More Than Just Right Answers<\/h2>\n

Imagine asking someone the time, and they confidently reply, \u201c2:30\u202fPM.\u201d They’re correct\u2014but what if their watch stopped exactly 12 hours ago, and they just happened to get lucky? This illustrates a classic philosophical point: the difference between a true belief<\/strong> and a justified true belief<\/strong><\/a>.<\/p>\n

And no, this isn\u2019t just a theoretical curiosity (yes, philosophy folks are sensitive about that critique). In practice, we do<\/em> care about how people form their beliefs\u2014and, by analogy, how AI agents arrive at their responses.<\/p>\n

In LLM system design, we evaluate not only if<\/strong> an agent is correct, but how<\/strong> it got there. A key metric here is tool correctness<\/strong><\/a>\u2014a measure of whether the AI used the right tools, in the right way, to arrive at its conclusion.<\/p>\n

This is where the analogy to justification comes in. Tool correctness isn\u2019t about whether the answer looks right, it\u2019s about whether the reasoning process was sound<\/strong>. It shifts focus from the content<\/strong> of the response to the process behind it<\/strong>.<\/p>\n

Enter Plan Validation<\/h2>\n

If you\u2019re not familiar with the Copilot Studio Kit<\/strong><\/a> (don\u2019t worry, we won\u2019t judge<\/em>), it\u2019s an open-source toolkit for testing and evaluating Copilot Studio agents. It allows you to define test cases, simulate user queries, and validate agent responses.<\/p>\n

One of the built-in test types is Generative Answers<\/strong>, which uses AI Builder<\/a> to LLM-judge whether the agent\u2019s response is semantically correct. That\u2019s important, mainly when you\u2019re evaluating the agent\u2019s knowledge capabilities<\/a>, like whether it retrieved relevant content or phrased an answer in a helpful way.<\/p>\n

But sometimes, semantic accuracy isn\u2019t enough.<\/p>\n

Some tasks don\u2019t produce meaningful responses at all\u2014like updating a database row or triggering a backend process. In those cases, there may be no content to evaluate<\/strong> (other than a vague \u201cdone\u201d?). Other times, the agent provides a fluent, seemingly correct answer\u2014but relies on the wrong tools<\/strong>, or skips tool usage entirely. The response looks right, but the resoning process behind it isn\u2019t reliable.<\/p>\n

Plan Validation is a recently added testing capability in the Copilot Studio Kit that focuses on tool correctness<\/strong>. Instead of evaluating what the agent says, it checks whether the expected tools<\/strong> were used during the plan.<\/p>\n

When defining a Plan Validation test, you specify:<\/p>\n