Configure your agent
Shape what your agent says and does — its voice and languages, what it knows, the flow of the call, and what happens when the call ends.
Open Voice Agents and pick an agent from the list on the left. Everything about it lives in four tabs.
| Tab | What you set there |
|---|---|
| Profile | Name, call type, voice, languages, guardrails |
| Context | Who counts as worth pursuing, knowledge the agent can draw on, and variables |
| Prompts | The conversation itself, step by step — what it says, what it captures, what it can do |
| Actions | What happens during and after the call |
Saving
Edits across every tab collect into one draft. A pill appears at the bottom of the screen when you have unsaved work — Save Changes commits all of it, Cancel throws it away.
The pill stays visible while you move between tabs. Change the voice on Profile, redraw a step on Prompts, and one Save Changes commits both.
Profile
Details
Agent name is what you call it. Company is who the agent says it's calling from.
Call type
Channel decides how this agent connects to people:
- Outbound — calls people: leads, customers, anyone on a list.
- Inbound — answers calls made to one of your numbers. Pick the Number to answer on; this is required and the agent can't be saved without it.
- Web — takes calls in the browser, with no phone line involved.
Two switches sit below:
- Detect voicemail — marks a call as voicemail when an answering machine picks up.
- Record calls — recordings and written transcripts are kept for every call.
Voice
Set the Model Tier, the Language, an optional Second language, and the Voice itself.
Stay in these languages controls whether the agent may switch mid-call:
- On — the agent only uses the two languages above.
- Off — the agent may switch to any language it supports to match the caller.
Turn Stay in these languages on if your agent drifts into a language you didn't choose. See Agent switches language.
Pronunciation
Say it like this teaches the agent to pronounce a name the way you say it — a brand, a product, a place.
Guardrails
Never say holds the things this agent must never say, promise, or imply — even if the caller asks directly.
Disclaimer
What the agent says about being an AI. Three switches: Say it's an AI if asked, Say it's an AI upfront (outbound calls only), and Say the call is recorded.
Leave them alone to follow the rules for your country — the tag under each one tells you what that rule is. Change one and this agent uses your setting instead; move it back to what your country says and it follows the rule again.
Context
Qualification
Decides who is worth pursuing, and in your own words rather than ours. It sits at the top of the tab because it changes what the call is for — everything below it is something the agent draws on.
Changes here reach the next call. The agent doesn't need rebuilding.
Leave it off and the agent still asks every question and records every answer — it just doesn't reach a verdict. Off is not the same as not asking.
Four steps, in order.
Decide who is worth pursuing
The switch that turns the whole thing on.
The questions, and what counts
Add question, then fill in:
- What should we ask? — write it plainly. The agent rephrases it to fit the conversation and the caller's language.
- How do they answer? — Text, Yes or no, One of a list, An amount of money, A number, or A date.
- Passes when — the test, and the value it's tested against: is more than, is at least, is, is not, is between, is one of, mentions, they answered, they never answered. Which ones you're offered depends on how the question is answered — you can't ask whether a yes/no is "more than" something.
- The options they pick from — only for One of a list, comma-separated. The agent names them only if the caller stalls.
- Save the answer as — filled in from your question. It only matters if you refer to the answer elsewhere.
- Must pass — off means the agent still asks and records it, it just doesn't count toward the decision.
A new question is added to the conversation when you save.
The decision
How the must-pass questions combine: All, Any, or At least a number of them. With two or more you can also group the last two as an either/or, for "this, and either of those".
How the result is labelled
Two words: one for a lead who clears it, one for a lead who doesn't. Your language, not ours — a recruiter shortlists, a lender pre-approves.
Renaming is safe. The rule and your past calls are untouched; only the wording changes.
A rule with no value set gates nothing. The agent asks the question and records the answer, but it counts toward no decision — which looks exactly like an agent that's working. The card warns you when a rule is in that state; set the missing value.
An unanswered question counts as not passed. Silence is never read as a yes.
Knowledge base
Documents this agent can look things up in while it's on a call. Tick one to attach it.
See Knowledge Base for how to add content.
Customer variables
Values filled in per call from your audience list in campaigns — a first name, a course, an appointment time.
Prompt variables
Fixed values the agent can drop into any prompt. Unlike customer variables, these are the same on every call.
Prompts
The Prompts tab shows the call as a flow: one card per step, wired together in the order the conversation can take.
Click a step to open its detail window. You can drag the window around and resize it, so the step you're reading about stays visible behind it. Draw a wire between two steps to say the conversation can go that way. Everything you change here saves with the rest of the agent.
The kinds of step
| Step | What you edit |
|---|---|
| Global Context | The agent's whole-call instructions — what it knows and how it behaves on every step |
| The opening step | Greeting message (spoken word for word) and Instructions |
| A conversation step | Instructions, Collect, and Tools |
| A closing step | Just the instructions — what the agent says as the call ends |
Instructions are guidance for the agent and are never read out to the caller. The greeting is the one thing spoken exactly as written.
Global Context starts out built automatically from the rest of the agent's settings. Type in it and you take over. Clear the box and it goes back to building itself.
Writing a prompt
Type @ anywhere in a prompt box to drop in a value — a customer variable, a prompt variable, or something an earlier step collected. Arrow keys pick, Enter inserts.
Start a line with # for a heading, - for a bullet, and wrap words in ** to bold them.
If a prompt refers to something nothing will fill in, the box says so while you're writing. Left alone, the agent reads the placeholder out loud on a live call.
Collect — what a step captures
The Collect tab lists what the agent must get before it can move on. Fields belong to a path out of the step, not the step itself, so a step with two ways out gets a tab per path with its field count on it.
Add field, then open it to set:
- Name — how you refer to it later. A field called
budgetis used in any prompt as{{slot_budget}}; the box shows you the exact wording as you type. - What kind of answer — text, a number, a yes/no, and so on.
- Must have this to move on — the agent asks again rather than continuing without it.
- Check the answer looks right — treats it as a known kind of value (a phone number, a date) and re-asks if it doesn't look like one. Pick the Kind of value; some kinds also let you pin an exact length.
- Ask again if this changes — if the caller corrects this later, anything captured after it is cleared and asked again.
Fields can be added but not removed here. A field the flow already references is a value your prompts and your tools may be reading, so taking one away is a bigger change than it looks.
Tools — what a step can do
The Tools tab is what makes a tool actually run. A tool set up under Agent Tools does nothing on a call until a step names it here.
Add Tool lists the ones that can run during a call. A tool that isn't ready is listed with the reason instead of hidden — Untested (it needs one test run first), Turned off, or a connection it still needs.
Open an attached tool and you get:
- When should the agent call this here? — leave it empty to follow the tool's own wording, or type here to change it for this step only.
- Its details, one box each. Leave a box empty and the agent asks the caller. Type a value and it's always sent without asking. Type a
{{variable}}and it's filled from the lead's details on campaign calls, with the agent asking when there are none. The line under each box tells you which of the three you're getting.
A detail the flow already collects under the same name is marked Collected — the agent uses what it captured, unless you've pinned a value over it.
A tool marked Every step stops being available everywhere the moment you add it to one step. From then on it only runs on the steps that name it.
The opening step has no Tools tab. Looking something up before the caller has even replied belongs to the conversation step after it.
If someone else changes this agent's flow while you have the canvas open, your unsaved flow changes can't be written on top of theirs — the save pill says Unsaved changes · not the flow, and Save Changes commits everything except the flow.
Your Profile, Context, and Actions edits are still safe to save. To get your flow changes back, reload the agent to pull in the newer flow, then redo them on top of what actually landed.
Actions
Call transfer
Hand the caller to a person when the agent can't help.
Set Transfer calls to, Give up after this many replies, and When to transfer — a plain description of the situations that should reach a human.
Not available for web agents, or on numbers that don't support it.
Email summary
Who gets emailed about this agent's calls.
Leads come from (Sources)
Which sources feed this agent. If nothing feeds it, this agent only calls people from campaigns and manual test calls.
Send call details to (Destinations)
Tick the destinations that should receive this agent's call details after every call.
A destination configured under Integrations does nothing until it is ticked here.
Test it
Test call at the top of the agent dials you so you can hear the agent for yourself.
For an inbound agent there's nothing to dial — call its own number instead, and the tooltip tells you which number that is.