Voice agents

Voice is a capability, not a type of agent. Give any agent a speaking voice and a phone number and it talks to people out loud, over the phone or in your app, keeping everything it already had: its prompt, tools, knowledge, memory, and transfers.

What makes an agent call-ready

First, a voice: either a realtime provider, or a classical speaking voice, picked on the agent's Voice tab. There is no default, so an agent without one has its calls rejected however they arrive.

Then a place to take the call. There are two, and an agent can use both at once:

• A phone number bound to the agent on Phone numbers. A number can carry its own voice override, which supplies the voice for that number even when the agent has none of its own.
• The website widget, which becomes a voice-call widget because the agent is set to Voice. Visitors call straight from your site, so this surface needs a widget key and no phone number at all. See the widget.

The Voice tab lists both surfaces and marks each one Call-ready, or names what it is still missing.

Autonomous Agents can have voice too. An agent that runs on a schedule and places outbound calls is a supported setup, so the type you pick never limits which surfaces the agent can use.

Two voice modes

An agent with voice runs in one of two modes, chosen per agent. Both support the same tools, knowledge, memory, transfers, and call features, so the choice is only about how the agent hears and speaks. You can move an agent between modes at any time in its settings.

• Realtime voice is the most natural option. The agent listens and speaks through one live AI connection, and turn-taking, interruptions, and emotional tone are handled inside the model as it talks. Choose it for the most conversational feel. See Realtime voice.
• Classic pipeline runs speech recognition, the language model, and speech synthesis as separate stages. Choose it when you want a specific synthesized voice, a specific recognition engine for a particular language, or separate control over each stage. Recognition is covered on the Speech recognition page; voices on Voices.

Voice & greeting

Pick the agent's speaking voice from the gallery, where you can play a preview of each one before you choose. Then set a voice greeting, the first thing a caller hears, and the agent takes it from there. Which voices you see depends on the families your workspace has configured. See Voices for the gallery and the voice families.

In-call controls

An agent with voice has controls that govern how a call behaves:

• End call lets the agent hang up when the conversation is done (end_call_enabled).
• Transfer hands off to a human or another line (call_transfer_enabled and call_transfer_number).
• Outbound allows the agent to place outbound calls (outbound_calls_enabled), with a cap on length (max_call_seconds).

How the agent handles quiet callers, interruptions, and turn-taking is covered in Voice tuning. What keeps a call smooth through a network hiccup is on Call reliability.

Inbound calls

To take inbound calls, bind a phone number to the agent. Callers who dial that number reach it, hear its greeting, and start a voice conversation. See Voice numbers.

Outbound calls

An agent can also place a call. Give it the agent and the number to dial, and the call's outcome and transcript appear as a voice conversation you can review afterward.