Voice agents¶
A voice agent answers a call. It is not a workflow: there is no canvas, no nodes, no graph. You write a prompt, choose a voice, and the conversation runs directly against a speech-to-speech model — the caller speaks, the model answers, and nothing sits between them.
That is the whole point. A conversation that waits for a graph to finish sounds like a conversation with someone reading from a script. A voice agent answers in roughly half a second because the reasoning is not on the path to the reply.
Voice agents are not voice in workflows¶
Both exist, and they are separate features:
| Voice agent | Voice inside a workflow | |
|---|---|---|
| Shape | a prompt and a voice | a graph with a transcribe node and an agent node that speaks |
| Turn | continuous, interruptible | one request, one answer |
| When | phone-like conversations | audio as an input or output of an automation |
Choosing one does not affect the other. See Node types for the workflow side.
What you configure¶
Agent — the system prompt and, optionally, a first message. Leave the first message empty and the agent waits for the caller to speak.
Voice — provider, model, voice and language. The trade-offs between providers are real and worth reading before you pick: see Providers.
Knowledge — knowledge bases are inlined into the prompt once, when the call starts. There is no retrieval mid-call, so there is no retrieval latency — but a long knowledge base costs money on every call and slows the first reply. The character counter warns you when the prompt gets expensive.
Analysis — up to two workflows that observe the conversation without shaping it. See Analysis hooks.
Calls¶
Every call is recorded: its transcript, how long it lasted, what it cost, and every analysis workflow it started. Open the Calls tab on the agent to read them back; each analysis run links straight into the normal execution viewer.
Cost is charged by wall-clock time at the model's per-minute price, shown next to the model when you pick it.
Calling the agent from your own product¶
The test panel in the editor is not the only way in. The same call is available over a public API: your backend mints a short-lived session token with a project API key, your frontend opens a WebSocket with it and streams audio. See Integrating a call.
There is no drop-in widget and no telephony yet — you build the client, and the protocol is documented so that you can.