Overview

A voice agent answers a phone call and holds a real conversation. It listens, decides what to say, calls out to your systems when it needs to, and reports how the call ended so your workflow can route on it.

You build agents under Automations → Agents in the left navigation. Calls reach an agent in one of two ways:

From a workflow. A workflow hands the call over using the AI Voice Agent node. Everything a workflow already does still applies: business-hours routing, screening, transferring afterwards. See Using an Agent in a Workflow.
Straight from a phone number. Assign one of your organization's numbers, including a number on a SIP trunk, to the agent on its Phone Numbers card, and inbound calls to that number go directly to the agent with no workflow in between. Nothing routes on a disposition afterwards, and the Set a workflow variable built-in tool isn't offered, because there's no workflow waiting.

The same agent can also answer in a chat widget on your website. Which of the two an agent is for is its own Channels setting, on the Persona card, and both are on by default. Settings that only mean something on a call are marked Voice only in the editor: speech and turn-taking, the idle timer, call limits, the telephony tools, keypad input, caller memory, phone-number routing, and post-call analysis. Switch Voice off and those settings are gone from the editor rather than marked, and the agent needs no voice. See Channels.

How a call flows through an agent

The call reaches the agent

An AI Voice Agent node passes the live call to the agent you selected, or the call came in on a number assigned to the agent.

The agent greets the caller

If you set a Greeting, the agent speaks it first. Otherwise it opens based on its system prompt.

The conversation runs

The caller speaks, the agent replies. Along the way it may search its knowledge base or call one of its tools, such as running a workflow or hitting an API, and speak the result.

The agent ends the call and reports a disposition

When the conversation reaches its natural end, the agent hangs up and reports one of the dispositions you defined. A workflow that handed the call over continues from the matching exit.

What you configure

The agent editor is a walk through the conversation, in the order one happens. A stepper across the top lists the steps; the page shows one step at a time, and the stepper is the only navigation. Steps are not locked in sequence: click any of them at any point, and Save Configuration in the header commits the whole agent from wherever you are.

Step	Cards on it
Identity	Persona
Routing & Memory	Phone Numbers, Chat Widgets, Returning Callers, Contacts
Conversation	Brain, Voice & Speech, Knowledge, Custom Tools, Built-in Tools, Keypad Input, Skills, Steering Rules, Dispositions, Safety
Data Collection	Data Collection
Wrap Up	Timeouts & Limits
Post-call	Post-Call Analysis
Review	Every setting above, on one screen

Some steps are dropped when they would hold nothing, so some agents walk four steps rather than seven. Routing & Memory goes for a chat-only agent in an organization that has neither chat widgets nor Contacts. Post-call goes whenever post-call analysis isn't enabled for your organization, or the agent has Voice switched off. Data Collection goes whenever data collection isn't enabled for your organization.

The stepper does not score progress — there are no ticks and no "3 of 5 done", because most of an agent is optional. The one thing it marks is the step holding the problem currently blocking Save Configuration, which it rings in red as you type rather than waiting for a failed save.

Review is the last step and the only place the whole agent is visible at once. It lists every card, including the ones you haven't filled in, with the values you've set; required fields that are still blank say so. Click a section heading to jump to the step that owns it.

Here is what each card holds:

Card	What it controls
Persona	Name, channels, system prompt, goal, greeting, and this agent's own variables
Phone Numbers	Which of your numbers route inbound calls straight to this agent
Chat Widgets	The website chat widgets that point at this agent
Returning Callers	Caller memory, the greeting returning callers hear, and reconnect after a dropped call
Contacts	Whether the agent recognizes its caller, and may update their record
Brain	Language model, reasoning, and conversation history
Voice & Speech	The voice it speaks in, whether callers can interrupt, and turn-taking
Knowledge	A knowledge base it can search for facts
Dispositions	The outcomes the agent can report when a conversation ends
Custom Tools	The actions you define: workflows, nodes, APIs, MCP servers, other agents
Built-in Tools	Call-control actions you switch on, such as transfers and keypad capture
Keypad Input	Whether a caller's keypresses reach the agent as an ordinary turn
Skills	Named procedures the agent applies when the situation calls for them
Steering Rules	Policies the agent's tool calls are checked against before they run
Safety	Content protections and caller-PII redaction
Timeouts & Limits	Idle behavior, call-duration limits, and failure handling
Data Collection	Details the agent asks for and records
Post-Call Analysis	What to pull out of the transcript once the call ends

Only Persona and Brain have to be filled in to get an agent working.

Some cards are a single row inside a group — Phone Numbers, Custom Tools, Built-in Tools, Keypad Input, Skills, Steering Rules and Dispositions. Each row carries a gear glyph; click it and the row's editor opens in a side panel. Chat Widgets is a row too, but a widget is its own record with its own editor, so it carries an open-in-new-window glyph and takes you to the widget instead of opening a panel. Brain, Voice & Speech, Timeouts & Limits, Data Collection and Post-Call Analysis keep their main settings on the card and put the rest behind a button in the card's header, which opens the same panel: Conversation History, Speech Tuning, Reliability, Data Collection Fields and Analysis Fields respectively.

A single-channel agent shows fewer of these. With Voice switched off the call-only cards are absent; with Chat off, Chat Widgets is. See Channels.

Changes save when you click Save Configuration, and take effect on the agent's next call. A call already in progress keeps the configuration it started with.

On an agent that uses versioning — which every agent created from now on does — Save Configuration writes to the draft you have open, and publishing that draft is what changes the configuration callers get. On the agent's published version you can still edit every field, but Save as New Draft replaces Save: your edits go to a new draft, and the published version stays as it was. The name is the exception, because it belongs to the agent rather than to a version, so a rename applies as soon as you save.

To hear a change before you save it, click Call Agent and talk to the agent in your browser. See Preview Calls.