Agent tests

Agent tests are simulated conversations: a fictional customer talks to your agent — with its script, its configured voice and its tools — and at the end it gets scored against the criteria you defined. Everything happens in writing and in seconds: no real call is placed, you spend no minutes and nobody receives a test call.

What a test is made of

• The persona: who the fictional customer is — how they speak, their mood, their context ("a customer three months behind on payments, in a hurry and distrustful").
• The scenario: what happens in the conversation and what that person wants to achieve.
• The success criteria: the list of things the agent MUST do ("offers a payment date", "identifies itself before asking for data"). Each criterion is scored separately.
• Optional — simulated responses: if your agent queries an external system, you can pin what it should return during the test.

Turn a call that went wrong into a test

1. 1
   Open the call detail

   Ask the Copilot for it (or find the call in Calls and ask it to "analyze call X").

2. 2
   Click "Create a test from this call"

   The button lives on the detail card, next to the recording.

3. 3
   Review what it proposes

   The Copilot reads the conversation and builds the persona, the scenario and the criteria — aimed exactly at making sure that failure does not repeat. Confirm when you agree.

4. 4
   Reuse it from now on

   The test is saved for that agent: run it every time you touch the script.

Run a test and read the result

1. 1
   Pick the test

   Ask the Copilot "show me this agent's tests".

2. 2
   Run it

   Say "run that test". Since it consumes AI, it asks for confirmation first.

3. 3
   Wait a few seconds

   The card updates itself while the conversation is simulated.

4. 4
   Read the verdict

   Passed or Not passed, how many criteria were met, which turn the conversation ended on and why it ended.

5. 5
   Dig into the detail

   Each criterion explains why it was considered met or not, quoting the conversation. And you can expand the full simulated transcript, including the tools the agent used.

The result is stored together with the script version that was live at that moment, so you can compare how your agent behaved before and after a change.

Good practices

• Make criteria observable: "offers a concrete payment date" can be verified; "is friendly" cannot.
• Three to six criteria per test: enough to cover the case, few enough to read the result at a glance.
• Build tests from the cases that hurt in production (the angry customer, the one who asks for a human, the one who does not understand) — those are the ones that break when the script changes.
• Run your tests before publishing a script change, not after.