Agent guide
Write effective checks
Give the agent a bounded job, a known starting state, and a result it can verify.
Check structure
Role + starting state + actions + expected result + important data
As a signed-in customer, start on /checkout, apply coupon SAVE10, complete payment with the test card, and verify the confirmation page shows the order number.Make it reliable
- Name the user role and starting page.
- Keep one business outcome per check.
- State exact values, roles, permissions, and test data.
- Prefer accessible labels, roles, and stable visible text in the product UI.
- State pass/fail evidence instead of asking the agent to ‘look around’.
Avoid
- Broad instructions such as ‘test the whole app’.
- Hidden setup that exists only in your head.
- Assertions based on timing, internal implementation, or unstable generated IDs.