How to test your agent in the sandbox

Use the sandbox to chat with your AI agent and validate replies, actions and tone before publishing it.

Jun 22, 2026

Watch this step in the video · The tester: chat, channel and contact

The sandbox is a safe environment where you can chat with your AI agent without any real customer seeing those conversations. It's the required step before pushing changes to production.

What the sandbox is for

  • Check the tone and quality of replies.
  • Confirm the AI uses the knowledge base correctly.
  • Verify it runs actions (booking, tagging, handoff) properly.
  • Catch invented or out-of-bounds answers.

How to start a test

  1. Open your AI agent and go to the sandbox tab.
  2. Type a message the way a real customer would.
  3. Read the reply and check whether it used correct data.
  4. Continue the conversation to simulate a full case.

Which cases to test

  • Frequent questions: prices, hours, location.
  • Edge cases: something the AI should not answer.
  • Actions: explicitly ask to book an appointment and verify it gets created.
  • Handoff: ask to talk to a human and confirm it escalates.

How to read what you see

  • If the AI invents data, information is missing from the instructions or knowledge base.
  • If the tone is off, adjust the instructions.
  • If an action fails, check that the tool is enabled and properly configured.

Iterate and test again

  1. Spot a concrete problem.
  2. Make a single change to instructions or knowledge base.
  3. Go back to the sandbox and repeat the same case.
  4. Once it's solid, then publish.

Final tip: keep a list of "test cases" you re-run on every important change. That way you make sure you don't break something that already worked while fixing something else.

Was this helpful?

Related articles

Message us on WhatsApp