Skip to main content
The voice agent is who speaks on the call. It defines what gets said, how it sounds, and how long the conversation can run. Agents live in the Agents tab of Voice. The table lists each agent with its name, its ID, and its creation date, and from there you open them to edit or create new ones with + Create agent.
Agents table showing the Name, ID, and Created columns, with the Create agent button at the top right

Voice agents listing

The agent ID (in the agent_... format) is what you use to reference it from the CLI and from campaigns. The table shows it ready to copy.

​
The editor

Opening an agent gives you four tabs and a Save changes button. Changes apply to subsequent calls, not to ones already in progress.
Agent editor with the Behavior, Tools, Knowledge base, and Analysis tabs, the system prompt, the first message, and the voice and language side panel

Agent editor on the Behavior tab

​
Behavior

​
System prompt

Defines the personality, the objective, and the rules the agent must follow during the call. It’s the equivalent of a Brain Studio AI Agent prompt, but written for a spoken conversation. Type {{variable}} to insert dynamic variables; the editor highlights them automatically. Beyond your own, you have system variables:
Including the date and time at the start of the prompt keeps the agent from using wrong time references — “yesterday”, “this month” — when the call happens at a different moment than the one you had in mind while writing it.
Since the prompt will be heard, not read, it pays to write it out loud: short sentences, no long lists the agent will recite in full, and a clear rule about what to do when the person interrupts or changes the subject.

​
First message

This is the first thing the agent will say when the conversation starts. Use it to introduce yourself, say who you’re calling on behalf of, and give the reason in one sentence. Right below it there’s an important setting: Block interruptions on the first message. When enabled, the user cannot interrupt the agent while it delivers its greeting.
Blocking guarantees the greeting is heard in full, but it can feel rigid if the greeting is long. If you need to say something mandatory — that the call is being recorded, for instance — enable it and keep it brief.

​
Voice and language

This panel defines how the agent sounds.
Panel with the Voice, Language, and Voice model selectors, and the Speed, Stability, Similarity, and Style controls

Voice and language panel

Always choose a voice that matches the configured language. A voice designed for another language produces unnatural accents and pronunciation.
For collections, support, or any case where precision matters more than warmth, raise stability and lower style: the agent will sound flatter but far more predictable. For sales or welcome calls, the opposite.

​
Model and call

Here you define which model reasons during the conversation and the call’s limits.
Panel with the Language model selector, the Maximum duration field in seconds, and the Time zone selector

Model and call panel

  • Language model: the model that interprets what the user says and decides the response.
  • Maximum duration (sec.): the call’s hard limit. When reached, the call is cut off regardless of where the conversation stands.
  • Time zone: the agent will interpret and express times in this zone.
The maximum duration protects your credit consumption against calls that stay open — for example when a voicemail keeps talking on the other end. That specific case can also be handled with the native voicemail detection tool.
The time zone matters more than it seems in multi-country operations: if the agent books an appointment or mentions business hours, it will do so in the zone you set here, not the caller’s.

​
Next step

With the agent configured, you need a phone number to receive and place calls.

Provision a number

Activate an ElevenLabs channel with SIP trunk or Twilio and assign this agent to it.

Give it tools

Enable telephony functions and connect the agent to your APIs.