Voice agents
Create, configure, test and publish the agents that answer your calls.
A voice agent is what talks to your callers. You describe its job, choose its voice and language, test it, and publish it. Only workspace owners and admins can open the Agents area.
Create an agent
- 1
Open Agents
Go to Agents in the sidebar. The All Agents list shows each agent's name, type, engine, voice, phone number and when it was last edited. The default agent has a Default badge.
- 2
Choose Create an Agent
Give the agent a Name and pick its type (see below). Start from Build from scratch or choose a template, then choose Create agent.
- 3
Or import one
Import accepts an agent file exported earlier. To export, open the row menu and choose Export.
Single prompt or conversational flow
Single prompt
Describe the whole task in one prompt and let the model run the conversation. The simplest way to start.
Conversational flow
Build the call as connected steps on a canvas, for calls that must follow a fixed order.
Set up a single-prompt agent
- Pick the Model, Voice and Language. A model without an API key shows (needs API key).
- Write the prompt. Type a backtick to insert a function and
{{to insert a dynamic variable. - Choose the Welcome Message: AI speaks first or User speaks first.
- Use Global Settings on the right for everything else.
Build a conversational flow
The canvas starts at a Begin node. Connect nodes with transitions, each with its own condition. Auto layout tidies the canvas. The node types are:
| Node | What it is for |
|---|---|
| Conversation | A step where the agent talks with the caller |
| Subagent | Hands a part of the call to a smaller agent |
| Function | Runs a function you have set up |
| Call Transfer | Transfers the call to a person or number |
| Press Digit | Waits for the caller to press a key |
| Logic Split | Chooses a path from a condition |
| Agent Transfer | Moves the call to another agent |
| In-Call SMS | Sends a text message during the call |
| Extract Variable | Pulls a value out of the conversation |
| Code | Runs a small piece of code |
| MCP | Calls a connected MCP tool |
| Ending | Ends the call |
| Note | A comment for your team; it does nothing in the call |
Global settings
Open Global Settings in the studio to configure how the agent behaves on every call. Its sections are:
- Agent Settings and Functions
- Knowledge Base & Memory: attach documents from your knowledge base
- Speech settings and Realtime transcription settings
- Call settings, Post call extraction and Post call memory settings
- Security & fallback settings, Webhook settings and MCPs
Call settings
| Setting | Allowed range |
|---|---|
| User Keypad Input Detection: Timeout | 0.5 to 15 |
| User Keypad Input Detection: Digit Limit | 1 to 50 |
| End Call on Silence | 10 to 1800 |
| Max Call Duration | 60 to 7200 |
| Ring Duration | 5 to 300 |
The same panel has switches for DNC Detection, Voicemail Detection, IOS/Android Call Screen Handling and IVR Hangup.
Call lifecycle
The Workflow tab on the left shows the steps around a call: dial in or out, pre-call functions, call started, call ended, post-call functions and post-call data extraction. Post-call extraction has a model picker that defaults to Workspace default.
Test your agent
- 1
Test in the studio
Choose Test to open the test panel. Use Test Audio to talk to the agent, or Test LLM to chat by typing. Fill in your dynamic variables first if the prompt uses any.
- 2
Test with a real phone call
Choose Phone Call, enter the number in Call to, and the agent calls that phone. Test calls are rate limited per hour.
Save versions and publish
- Your edits stay a Draft until you choose Publish. Once live, the button reads Update.
- Save draft keeps your work without changing live calls.
- The Versions panel lets you Save version, make a version live (new calls use it straight away), Restore to draft, or Delete version (not available for the live one).
- Configure Tags labels versions. The tag prod marks the live version, and you can add your own such as staging or qa.