Build
Testing your agent
Test Audio, Test Chat, phone tests and simulations.
Four ways to test
VoiceSquad gives you four test surfaces, and they are not interchangeable. Pick by what you are trying to learn.
| Surface | Where | Use it for | Real call |
|---|---|---|---|
| Test Audio | Editor right panel | Hearing the agent. Voice, pacing, interruptions, does it actually sound right. | No |
| Test Chat | Editor right panel | Fast logic checks over text, when you do not need to listen. | No |
| Test Phone Call | Simulation dropdown | End to end proof over the real carrier path, on a real handset. | Yes, billed |
| Text Simulation | Simulation dropdown | A saved library of cases you run again before every publish. | No |
Test Audio
The right panel of the editor opens on Test Audio. Press Test your agent, allow the browser microphone, and you are talking to the draft in front of you. The panel moves through connecting, then Live with a pulsing dot and a duration timer, then Call Ended, where New Test Call starts a fresh one. The agent needs to be saved and to have a non empty prompt before the button works.
When the agent invokes an action mid conversation, the transcript shows a tool call card with the tool name, its status, the arguments it sent and the result it got back. That card is usually where a confusing answer turns out to be a wrong argument rather than a wrong prompt.
Test Chat
The Test Chat tab runs the same agent config over text. Type into Type your message... and the agent answers in the same transcript view, tool call cards included. New Chat clears it and starts over. It has the same requirement of a saved prompt, and it honours your welcome message settings, so an agent set to speak first greets you here too. Use chat when you are iterating on wording and logic, and audio when you care how it lands in the ear.
Test Config
The gear icon next to the Test Audio and Test Chat tabs opens the test config, which controls what the next test run pretends the world looks like. Two things live there: It opens pre-filled with your payload schema's example values, or plausible generated ones, so a run is ready to start as soon as you open it.
- The values your dynamic variables resolve to for this session, so you can rehearse the same call with different inputs without editing the agent.
- Mocked results for specific tool calls, so you can rehearse the unhappy paths. Force the lookup to return nothing, or the booking to fail, and hear how the agent copes without breaking anything in the real system behind the tool.
Test Phone Call
Under the Simulation dropdown in the header, Test Phone Call places a genuine outbound call through exactly the path your own systems would use. It is available only once the agent is published and configured for outgoing calls, and until then the item is locked with a tooltip explaining which condition is missing.
- Publish the agent, with the call direction set to Outgoing.
- Open Simulation, then Test Phone Call.
- Enter the destination under Call to, for example
+14155551234, and fill one input per required payload field. - Press Place Call. The modal confirms the call was dispatched and links you to Conversations to follow it.
Text Simulation
Everything above is a one off run that leaves nothing behind. Text Simulation, the bottom sheet under the same Simulation dropdown, is the agent's durable test case library: the set of scenarios you commit to checking before every publish.
- Build the library. Write cases by hand, use the generate action for one at a time, or generate a batch from the prompt.
- Press Run Simulation. Simulation runs are metered, so a cost confirmation appears before anything starts. Confirm to run the batch.
- Read the results. Each case row carries its name, Last Tested, Status and Reason.
- Open a case for the detail, which has tabs for Test Details, Conversation Transcription and Evaluation. The evaluation uses the same scoring you see on real calls in Conversations.
The value compounds. Every time a real call goes wrong, add it as a case. Within a few weeks the library encodes what your agent must never get wrong again, and a publish stops being a leap of faith.
Version history makes iteration safe
None of this touches production. Editing a live agent clones it into a draft, and only that draft is what the playground and the simulation run against. If a published version turns out to be worse than the one before it, the old version is still in History: open it, press Edit to clone it back into a draft, and publish. See agents and versions for how drafts, published and older versions relate, and the agent editor for the settings you are testing.