Agent settings

The agent builder groups its sections in three parts: Build, More settings (folded until you open it) and Ship. This page lists every setting in the order the app shows it, with its default for a new agent. Templates, Create with AI and the Start from scratch wizard fill in many of these for you.

You need a workspace role of Builder or higher to change settings. Viewers can open every section except Test, read only: the top bar says Read only.

All changes save to the draft on their own. Real calls only use them after you publish.

Top bar

Item	What it shows
Save state	Saving…, Saved, Not saved · Retry, or Someone else saved · Reload when another person changed the draft.
Test this change	Appears after you change the draft. Starts a test call (Ctrl+Enter), or a chat from its menu.
Price chip	The price per minute and typical reply time of the current models, for example ₹3.42/min · replies in about 1.1 s. Click it for the breakdown.
Lines	Which version the test line and the live line run, and whether real callers can reach the agent.
Publish	Publishes the draft as a new version.

The Test panel on the right always uses the draft. See Test an agent in your browser.

Build

Basic info

Setting	What it does	Default
Name	The agent’s name, up to 120 characters.	The name you gave it.
What it does	Appointment booking, Lead capture, General assistant or Custom form. Decides the built-in tools and what the agent saves. See Collect details from callers.	Set when you create the agent. From scratch: General assistant.
Persona	One line on who the agent is, up to 400 characters. It shapes how the agent speaks. Refine rewrites it with AI.	Set when you create the agent.
Language	Multilingual (auto-detect), English (India), English (US), English (UK), English or Hindi.	English (India), or English (US) for organisations in the United States.
Tone	Friendly, Professional, Casual or Formal.	Friendly
Follow the caller’s language	Switches language when the caller does, for example from Hindi to English.	Off
Timezone	The agent’s clock: today’s date and booking times.	Your organisation’s timezone.
Maximum call length (minutes)	1 to 120. The agent wraps up politely before the limit. Also in Voice tuning, in seconds.	10
Tags	Press Enter to add. Used for search and filters.	None

Prompt

Setting	What it does	Default
Instructions	The agent’s job, its rules and how to speak. Up to 20,000 characters.	Set when you create the agent.
Greeting	The first thing the agent says. Up to 2,000 characters. Empty: the agent greets in its own words.	Set when you create the agent.
Closing line	Said before the agent ends the call. Up to 2,000 characters.	Empty
This agent can	Read only. The built-in tools for the agent’s type, and your custom tools.
Variables	Your own {{variables}}, with Add variable. Also lists the built-in ones.	None

Each text field has Insert variable and Refine. A variable has a Name, Shown as (optional), a Type (Text, Number, Date, Phone number, Email or Yes / no), Usually comes from, Default value (optional), Say instead, if unknown, Description (optional) and Needed when the call starts. See Write and edit the agent’s instructions.

Pathway

Only for pathway agents, and only when pathways are on for your organisation: the canvas of steps the call follows. See Design a multi-step call with a pathway.

Data to collect

Only for single-prompt agents. The agent asks for these fields when What it does is Custom form.

Setting	What it does	Default
Question / label	What to ask for, up to 80 characters.
Kind	Text, Number, Rating, Email, Phone, Date, Date & time, Choice or Yes / no.	Text
Key	How the value is stored. Fills in from the label.
Options (one per line)	For Choice: at least 2.
Min, Max	For Number and Rating.	Rating: 1 and 5
Required	The agent must get this one before it reads back.	On
Prefix (under Record ids)	2 to 5 capital letters for saved record ids. Custom forms only: bookings always use APT, and leads LEAD.	REC

Up to 15 fields. Appointment booking agents also have Send bookings to:

Setting	What it does	Default
Tool	Also send each confirmed booking to one of your tools that runs during the call, or to a Salesforce event (connected org).	Don’t send anywhere else
Appointment length (minutes)	5 to 480.	20
Request body	Plain JSON (webhook), or Salesforce flow ({"inputs":[…]}) for a Salesforce flow or invocable action.	Plain JSON (webhook)
Arguments	A value for each of the tool’s parameters: fixed text or a {{booking.…}} placeholder.

Voice & models

Setting	What it does	Default
Presets	Budget, Balanced, Performance or Ultra. Each sets every model at once.	Balanced
How the agent hears and speaks	Listen · think · speak (three models) or All in one (one speech-to-speech model).	Listen · think · speak
Listening	The speech-to-text model.	Deepgram Nova-3
Thinking	The LLM.	GPT-4.1 mini
Speaking	The text-to-speech model.	ElevenLabs v4 Turbo
All-in-one model	With All in one only.	OpenAI Realtime mini (with Ultra)
Words to listen for	Names speech-to-text should hear right. Up to 50, 40 characters each. Only Deepgram Nova-3 uses them.	None
Voice	How the agent sounds. ▶ plays a sample; Use my greeting plays your greeting.	The Balanced preset’s voice.

If your platform can’t reach one of a preset’s models, the preset uses another one, and the pickers show what it chose. See Choose the voice and language.

Knowledge

Shown when knowledge bases are on for your organisation. Switch on each knowledge base the agent may search during a call. None are on for a new agent. See Answer questions from your documents.

More settings

Tools

Up to 20 per agent. See Give the agent tools.

Setting	What it does	Default
Name	Lowercase verb_noun, for example get_order. The AI sees it.
When to use it	The AI reads this to decide when to call the tool. Required for tools that run during the call.
Runs	During the call, Before the call or After the call.	During the call
On / Off	Turns the tool off without removing it.	On
Method	GET, POST, PUT, PATCH or DELETE.	GET (from curl: as pasted)
URL	https:// only. Can hold {{variables}}.
Timeout (ms)	500 to 60,000. Above 8,000 for a tool during the call, the editor warns you.	5,000
Body	For POST, PUT and PATCH: JSON or Form (urlencoded).	JSON
Headers, Query string	Each value is Text or a stored Credential.	None
Parameters	What the AI fills in: name, type, Required, description.	None
Save from the response	A variable name and a JSON path, for example $.order.status.	None
Test	Run test calls the tool for real with Arguments (JSON).

Salesforce

Needs your organisation’s Salesforce connection: an Owner or Admin connects it in Integrations. See Connect Salesforce.

Setting	What it does	Default
Use Salesforce on this agent’s calls	Turns the rest on.	Off
Look up the caller (while the phone rings)	Finds the caller by their 360 CTI record, else their number (Contacts and Leads), so the agent knows their name, company and open cases.	On
Create a Case, Create a Lead, Create a follow-up Task, Create an Event (during the call)	Each gives the agent a tool to create that record. Add field maps its answers to Salesforce fields.	Off
Log the call as a Task (after the call)	Adds a completed Task with the outcome, summary and duration to the caller’s record.	On

Post-call analysis

Always included: Summary, Outcome, Sentiment and Transfer reason. The outcome list is your organisation’s, in Settings → Outcomes. Your own fields, up to 20:

Setting	What it does	Default
Name	What the field is called in call details.
Type	Yes / no, Text, Number or One of a list.	Yes / no
Key	How the value is stored. Fills in from the name.
What to look for	How to decide the value. The AI reads it.
Options (one per line)	For One of a list.

See Set up post-call analysis.

Guardrails

Setting	What it does	Default
Check every caller turn	Turns all checks on or off.	On
Caller wants a person	Switch, confidence, action.	On · 80% · Transfer to a person
Caller asks not to be called		On · 80% · Stop calling and end
Abusive language		On · 85% · Warn once
Off topic		Off · 85% · Record only
Urgent safety or medical matter		On · 70% · Transfer to a person
Custom checks	Up to 5: Label, Check, confidence, Then, Line to say, On.	None (new ones start at 80%)
Transfer to	A phone number in your organisation’s country, or a CTI queue.	Empty

Confidence is set from 50% to 99%. See Add guardrails.

Voice tuning

Everything that changes how a call feels. The defaults suit most calls.

Interruptions and turn-taking

Setting	What it does	Default
Interruption style	Patient (lets callers pause and think), Balanced (natural for most calls) or Responsive (replies fast and stops the moment the caller talks).	Balanced
Shortest wait after the caller stops	Replies sooner when the caller clearly finished. 0 to 5,000 ms.	500 ms
Longest wait	Most time it gives a caller who paused mid-thought. 0 to 10,000 ms.	1,800 ms
Ignore “uh-huh” while speaking	Short words like hmm, okay, haan ji don’t stop the agent.	On
Carry on after a false start	If a cough or noise cut the agent off and the caller says nothing, it picks up where it stopped.	On
Start replying early	Prepares the answer while the caller finishes.	On

Patient waits 800 to 3,000 ms, and Responsive 300 to 1,200 ms. Changing a wait yourself shows Custom waits; picking a style again resets them.

Silence and call length

Setting	What it does	Default
Check in when the caller goes quiet	Asks once, in the caller’s language, before hanging up. After: 2 to 60 seconds. It says: the line.	On · 10 seconds · Are you still there?
Hang up after total silence of	The agent says a short goodbye first. 5 to 300 seconds.	30 seconds
Maximum call length	The same setting as in Basic info, in seconds: 30 to 7,200.	600 seconds

Sound

Setting	What it does	Default
Noise removal	Off, Background noise (removes traffic, fans and chatter, no extra cost) or Voice isolation (keeps only the caller’s voice, billed per minute).	Background noise
Background ambience	A faint room sound under the agent’s voice: None, Quiet office, Call centre, City street or Keyboard typing.	None
Thinking sounds	Soft typing while the agent looks something up.	Off

Model behaviour

Setting	What it does	Default
Use the built-in speaking style	Adds a short guide for natural phone replies. See the built-in speaking style.	On
Creativity (temperature)	0 to 1.5. Lower is more consistent and on-script. Not shown for All in one models.	0.6
Backup language model	Takes over mid-call if the main model fails. Automatic picks one from another provider.	Automatic

Show raw values lists the exact values the voice agent uses, including a few that only the API changes: custom interruption settings, the end-of-turn model, how soon a hold line plays (1 second) and how soon a stuck model switches to the backup (2.5 seconds).

Ship

Section	What it’s for
Test	Test calls and chats, simulated callers, scenarios and results per version. Needs Builder or higher. See Test an agent in your browser.
Versions	Every published version, comparisons, rolling back and the draft history. See Publish and roll back versions.
Metrics	This agent’s phone calls in the last 7 or 30 days, compared with the period before.

Limits

For the sizes and counts the app enforces across the workspace, see Limits.