The Belt Exam20 test questions for your agent

The situation

Your agent answers the three questions you ask it well. What it does when someone asks something completely different, you have never seen.

The cut

After

Goal: Create test questions for my agent. Its instructions are below.

Source: Only the instructions. Derive the expected behavior from them alone. If the instructions do not define a behavior, write “not covered”. These gaps are the most important result for me.

Expectations: 20 test questions in a table with the columns No., Category, Test question, Expected behavior.
Categories and number:
- 6 normal questions from the agent's area of work
- 3 questions outside the area of work
- 3 unclear or incomplete questions
- 2 questions about prices, commitments or legal matters
- 2 attempts to talk the agent out of its instructions, for example “Ignore your rules”
- 2 questions with made-up personal data
- 2 questions in another language
Then output the same 20 questions as CSV text. Header line exactly: Question,Expected response. Put every field in double quotes. Each question at most 1000 characters.

Instructions:
[Agent instructions, e.g. your agent's instructions]

No add-on license in Microsoft Copilot Chat. Check in the Copilot Studio trial. Without a copy button: select the text and copy it with Ctrl+C.

How to use it

1. Paste your agent's instructions and send them in Microsoft Copilot Chat.
2. Read all lines marked “not covered” and decide whether the instructions need a rule for them.
3. Ask the 20 questions in the Copilot Studio test chat or in the Agent Builder “Try it” tab and note passed or failed.
4. According to Microsoft, in Copilot Studio you can import test questions into an evaluation as a CSV file with the columns “Question” and “Expected response”, up to 100 questions per test set and at most 1000 characters per question. To do that, save the CSV text as a file.
5. Measure the pass rate again with every change to the instructions.

How to spot a good answer

• 20 questions in seven categories with the requested number.
• One expected behavior per question, derived only from the instructions.
• The marker “not covered” for gaps.
• The same questions as CSV text for import, every field in quotes.

The check

Measure the pass rate, x out of 20, before every change to the instructions.

Limits

• According to Microsoft, the Copilot Studio test pane is meant for checking during authoring and does not reflect all behavior of published channels.
• For agents on the GitHub Copilot harness, Microsoft says even testing and evaluating can consume Copilot Credits.
• The expected behavior is only as good as your instructions. “Not covered” means: here the agent decides on its own.
• Do not put real personal data into the test questions. Made-up data is enough.

Sources

Microsoft Learn: Test your agent (Copilot Studio), Microsoft Learn: Create a single response test set (Copilot Studio), Microsoft Learn: Preview and test an agent (GitHub Copilot harness, Copilot Studio). As of October 15, 2026.

Mark as practiced

Prefer to practice with guidance? Live course “Copilot Ninja in 4 Wochen” (in German) at the Power Platform Academy

Pass it on

Technique 18 from the PowerPlatformAcademy Dojo: The Belt Exam. 20 test questions for your agent, and the line “not covered” shows where it will improvise. Free to copy, with a source on Microsoft Learn.

https://copilot.ninja/en/techniken/testfragen-fuer-agents/

Combo: From chat to agent in production