Tests & comparing versions

Open Build → Tests.

Write a test

A test is a customer and what the agent must (or must never) do.

• The customer — either fixed messages sent in order, or AI plays the customer: describe who they are and what they want, and an AI writes their messages for up to 8 turns.
• Checks — *must use* / *must never use* an action or built-in tool (hand to the team, save a lead, verify email, start a procedure…), *hand to the team* yes/no, words replies must or must never contain, and a requirement in your words checked by AI (e.g. *asks for the order number before promising anything*).
• Advanced — the customer can count as already verified; give actions sample results (e.g. an order that is lost); optionally run read-only actions for real.

Not sure what to test? Suggest tests writes ideas from your business overview, actions and real customer questions. In the inbox, Save as test turns any real chat into a test.

Run them

Run one test, or Run all after you change instructions, the model, sources or actions. Each run shows every check, the full conversation and which tools were used. Earlier runs are kept so you can see when something broke.

Nothing real happens

Tests run in a sandbox: actions that change data never run (they return a test result), handoffs, leads, approvals, emails, alerts and webhooks are only recorded, and the test chat is deleted afterwards. Each plan includes free test replies every month; after that, tests use normal credits.

Compare versions

Compare versions sends the same customer messages to two or three versions of your agent side by side — different models, creativity or instructions — so you can see which answers better before you change anything. Pick from real questions your customers asked. Uses credits like a normal chat; nothing is saved.