Testing
Try your agent with test calls and chats, and score it automatically with simulated customers
Test your agent before customers meet it, and again each time you change its instructions. There are three ways, from quick to thorough.
Test calls
Press Test call in the dashboard and talk to your agent from the browser, exactly as a customer would. You hear its voice, see the live transcript and every tool it runs, and can type a message instead of speaking.
Test calls use your plan's free test-call minutes first: 3 to 15 a day, depending on your agent plan. After that, they come out of your call minutes. A test call is stored like any other conversation, on the test channel.
Test chats
Type to the agent as a WhatsApp customer would. The agent answers with the same instructions, knowledge and tools as on WhatsApp. Test chats are free, up to 120 messages an hour.
Tools are simulated: the agent is told each tool succeeded, and no request reaches your systems. Turn on Run tools for real to call them.
Automated tests
On Tests, a simulated customer talks to your agent while an AI reviewer reads the conversation. The reviewer checks each of your success criteria, then the scenario passes or fails.
A scenario has:
| Field | Example |
|---|---|
| Scenario name | Angry customer about a late order |
| Customer persona | A man from Algiers in a hurry who speaks Darija. His order is five days late and he wants a refund. |
| Customer's goal | Find out where the order is, or cancel it |
| Success criteria (up to 8) | The agent apologises and offers a clear solution · It never promises a delivery date it doesn't know · It offers a callback from the team |
| Turns | 2 to 10 exchanges (6 by default) |
The simulated customer speaks your agent's language, plays the persona, and reacts to what the agent actually says. The agent runs with its real instructions and knowledge. Its tools never reach your systems: it's told they succeeded with the values it chose.
Suggest scenarios with AI writes three scenarios from your instructions (an easy one, a tough customer and an edge case), which you can edit. You can keep up to 30 scenarios per agent and Run all after every change.
Each run counts as 5 messages from your agent plan. The result keeps the full conversation, the verdict for each criterion with its reason, and a short review.
Write criteria you can check from the transcript alone: "the agent asks for the wilaya before giving the delivery fee" is testable, "the agent is helpful" is not.
What to test
- The happy path: a customer who wants exactly what your agent is built for.
- Missing knowledge: a question your knowledge base doesn't answer. The agent should say so and offer a follow-up, never invent an answer.
- The hard customer: angry, rushed, confused, or asking for a person.
- Off-topic and pressure: a request for a discount you don't offer, or for the agent's instructions.
- The language: a caller who switches to French or English. The agent should keep to its own language.
Was this page helpful?