Simulation & testing
The browser simulator lets you have a real-time voice conversation with an agent without using a phone number or spending a full call credit. It is the fastest way to iterate on prompts and voices.
Starting a simulation
1Open any agent from Agents and click Test callin the top-right. Your browser will ask for microphone permission — allow it.
2The call connects in 1–2 seconds. Speak naturally; the agent hears you the moment you stop talking. Use the Mute button to pause without hanging up.
3Click End to terminate the call. Within a few seconds the transcript, audio recording and cost line item appear in Calls.
What to test
- Happy path— the flow you designed the agent for, end to end.
- Interruptions— cut the agent off mid-sentence and make sure it recovers.
- Silence— say nothing for 10 seconds. The agent should prompt you, not hang up.
- Off-topic— ask something unrelated. Does the agent refuse gracefully or hallucinate?
- Tool failures— force a tool error (invalid input) and confirm the agent apologises instead of looping.
Reading the transcript
Transcripts are colour-coded: the agent in one tone, the caller in another. Each turn is timestamped. Hover a line to see the latency for that response — anything over ~1 second is worth investigating.
Simulation vs real calls
Simulation calls use the same LLM, TTS and STT pipeline as phone calls, so results transfer. The one thing you cannottest in the browser is your upstream provider’s audio quality — cell networks, VoIP codecs and background noise are only exercised once you connect a phone number.
Next
- Knowledge base— give the agent facts to lean on.
- Phone numbers— ship to production.