Skip to main content
An agent can sound confident while giving an incomplete or incorrect answer. Test the job you expect it to do, the questions it should not answer, and the action or handoff a visitor should receive. Start with a small set of high-value cases and run them again after changes to instructions, knowledge, or actions.

Build a test set

For a fictional order-management agent, a useful first set might be: Use examples from your own product and approved test data. For each case, define what a good result contains and what would make it misleading. Avoid requiring one exact sentence when several accurate answers would help the visitor.

Save a repeatable case

On the agent’s Tests page, select Add test → Create new test. Choose a test type that matches what you want to check:
  • Single answer test checks a response against your description of a good answer. You can add good and bad examples to clarify the distinction.
  • Full conversation test uses a simulated visitor and success criteria to exercise a longer exchange.
  • Action test checks behavior involving an action.
Name the case so the team can understand its purpose later. Add the relevant conversation context and describe the expected behavior, then save or choose Create & run. For an answer test, specify the facts and boundary that matter; for example, “Explain the late-order filter, do not claim access to a specific customer’s account, and offer support for account-specific questions.” Use Run all or select cases and run the selection. Review the displayed credit estimate and any available run options before starting. Open results to inspect the actual conversation and checks, especially for failures or unexpected passes. A passing result is evidence about that case, not a guarantee about all visitor questions.

Test the actual channel

Saved cases complement, but do not replace, a live check of the published experience. Open the website installation or share link in a separate session and try the greeting, a normal question, an out-of-scope question, and the intended handoff. Verify that the visitor can use the surfaced demo or link.
Actions in Try it can run for real, including bookings, CRM updates, and API calls. Review connections and use safe test data before testing. See Build and launch a demo agent.
When a case fails, identify whether the issue is the source, instructions, action setup, or the test’s own expectation. Change the relevant draft, repeat the case, and review the other important cases before publishing.