Platform
An agent is one configuration — a prompt, a voice, a set of tools, a phone number — that can answer a chat message, a browser call or a real phone call identically, because all three read the same settings this page walks through.
Open Agents in the sidebar and select New agent. You land on a blank agent with six tabs across the top — Overview, Models, Telephony, Voice, Tools and Sandbox — and this guide covers each in that order. Nothing is created until you save; try settings freely.
Name is for your own dashboard, never spoken. The System prompt is what the agent is told about who it is and how to behave — the single biggest lever over how a call goes. The Welcome message (labelled "Opening line used at call start") is spoken first, before the caller says anything.
Three sub-tabs, one per stage of a spoken turn: what the agent hears, what it decides, and how it answers.
Transcription provider picks the engine that turns the caller's speech into text. Language — "the language the agent listens and speaks in" — offers Auto detect or a specific language, filtered to only the languages the provider you picked actually supports. A language hint is passed to the provider without forcing it: it is a preference, not a hard lock, so the agent still copes if a caller switches language mid-call.
Language model is the model this agent thinks with. Thinking level (Minimal, Low, Medium, High) trades latency for how much a model reasons before answering, where the model supports it at all. Max reply length caps how long a single answer can run, and Temperature controls how much variety is in its wording — leave either on Model default unless you have a specific reason not to.
Pick the voice this agent speaks with, and its language, the same picker and the same voices as the Model Playground — try a voice there first if you are not sure which one fits.
Set the direction this agent handles — inbound, outbound, or both — and which telephony account and number it uses. This is covered in full, including bringing your own carrier account, in the Telephony guide.
The Voice tab is about the mechanics of a spoken turn — when the agent may be cut off, and how long it waits for the caller to keep talking — not the voice itself.
Allow interruptions switches the whole feature on or off. Interrupt on chooses the trigger: Spoken words, which waits for a minimum number of words before cutting the agent off (so a stray cough or a single "yeah" does not interrupt it), or voice activity, tuned by VAD confidence, Start delay and Stop delay (in seconds), and Minimum volume — "audio quieter than this is ignored," which keeps background noise on the line from triggering an interruption.
End-of-turn pause is how long the agent waits in silence before deciding the caller has actually finished talking — "speech must last this long to count as the caller talking" is the matching rule on the other side, filtering out a brief noise that is not really speech.
Ask the caller if they are still there turns on a check-in after a stretch of silence; set what it says (a Fixed phrase or left to the model) and how many times it checks before giving up. Under Safety limits, choose Immediate or Graceful for how a call ends when the caller disconnects — immediate stops the moment the line drops, graceful lets the agent finish the sentence it was already speaking.
Every tool you have built in the Tools Library is listed here, grouped by kind; tick a tool on to give this agent access to it. Only what you tick here reaches the agent — nothing in your Library is available to an agent just because it exists, and an API tool that has not passed its verification test cannot be ticked on at all.
Two built-in tools live in their own Call handling group, since they need no setup in the Library first: Voicemail detection (outbound calls only) leaves a message when the call reaches voicemail — write one, or leave it blank to let the agent compose one — and Call transfer hands the call to a human number you set. Both take the same "when to use it" and "say while it runs" fields as every other tool.
Test the agent exactly as saved — same prompt, same tools, same voice tuning — before anyone calls it for real.
Type a message and read the agent's reply as text, no audio involved. Useful for iterating on the prompt quickly, and it still runs the agent's real tools — a tool called while testing in chat shows up in Call history the same as one called on a live phone call.
Talk to the agent through your computer's microphone and speakers, over WebRTC — the closest thing to a real call without dialling a number, and it carries every setting on this page: the same tools, the same voice tuning, the same interruption behaviour a caller would get.
Enter a number with its country code and select Call my phone; the agent places a real outbound call to it, through the telephony account set on the Telephony tab. This is the only sandbox mode that spends telephony minutes, billed the same way any other outbound call is — see How telephony minutes are billed.
An Owner or Admin can create, edit and delete any agent in the workspace. See Team & Workspace for what a Member can see instead.