Create an agent

An agent is the voice worker that actually talks to your callers. Think of it as a really well-briefed teammate: you tell it who it is, what the call is about, what it’s allowed to do, and how to behave - once. After that it can take thousands of calls a day, in the language your customer speaks, without getting tired or going off-script.

You build an agent once and reuse it everywhere: point an outbound campaign at it to make calls, or point a phone number at it to answer calls. Nothing an agent does goes live until you publish it, so you can take your time, save a draft, and come back.

The builder is a short 3-step wizard: Basics & voice → Call context → Tools. We’ll walk through every single field, what it does, and how to fill it in.

Before you start: you don’t have to get everything perfect. Save a draft at any point (top-right), test it, and tweak. The only step that matters for going live is Publish at the end.

Step 1 - Basics & voice

This step decides who the agent is and what the caller hears. The voice and language are literally the first thing a customer experiences on the call, so they set the tone for everything that follows.

Name

A label for you and your team - the caller never hears it. It’s how you’ll find this agent later in your list.

Tip: name it after the job it does, not the campaign it’s for. “Lead Qualifier” is still meaningful next year; “Diwali Push 2026” isn’t.

Agent type

How the agent runs the call. Two choices:

Conversational - a normal back-and-forth call driven by a persona. It listens, answers, asks questions, handles objections. This is what you want for sales, qualification, support, follow-ups - almost everything.
Announcement - the agent delivers one fixed message (a reminder, a notice), lets the caller acknowledge it, then politely asks to end the call. It won’t hold a real conversation. Use it for “your appointment is tomorrow at 4pm” style notifications.

Vertical

The industry this agent works in (lending, healthcare, e-commerce, and so on). It’s not just a label - it helps the platform recommend the right compliance pack and flags whether the call is regulated.

Regulated

A toggle that says “this call is in a regulated space” (like lending or collections). When it’s on, the platform turns on extra legal safeguards automatically - for example, verifying it’s speaking to the right person before revealing any account details. If your vertical is regulated, leave this on.

In plain terms: this is a safety switch. When in doubt in a regulated industry, keep it on - it protects you.

Reply model

Which AI brain generates the agent’s replies. Three options:

Platform default - we pick a sensible, fast model for you. Leave it here unless you have a specific reason to change it.
Fast (Haiku) - lowest latency; the agent replies quicker. Great for simple, high-volume calls.
Accurate (Sonnet) - a bit slower, but reasons more carefully. Worth it for complex or high-stakes conversations.
Tip: don’t touch this until you actually have a complaint to point at (“it’s too slow” → Fast; “it’s making mistakes on tricky questions” → Accurate). Otherwise, default.

Voice

The speaker your caller hears. You can preview each one. Pick a voice that fits your brand and audience - a warm, everyday voice usually beats a formal one on the phone. See Voice & language to compare speakers.

Language

The main language the agent speaks in - Hindi, Hinglish, English, and several more.

Set the language your list actually speaks, not the one your team writes in. If your customers speak Hindi, the agent should open in Hindi.
You can also set a secondary language - the agent will detect and switch mid-call if the caller replies in it. Only add one if your list is genuinely mixed; switching costs about a fifth of a second each time, so there’s no point adding it “just in case.”

A good Step 1 looks like: Lead Qualifier · English (IN) + हिन्दी · Shubh (voice).

Step 2 - Call context

This is the most important step. Everything the agent says is generated from what you write here. A vague brief produces a vague call, and no amount of tuning later will fix a fuzzy purpose. Write it like you’re briefing a new hire who’s about to pick up the phone.

You’ll notice a “Draft from brief” button - describe the call in a sentence or two and the AI fills these fields in for you as a starting point. You can then edit everything.

Company / organization name

Who the agent is calling on behalf of. The agent uses this to introduce itself (“Hi, I’m calling from …”). On a new agent this is pre-filled with your workspace’s company name, but you can change it.

Purpose

The single most important field: why is this call happening? Write it as one clear sentence someone could act on.

Good: “Remind the customer their EMI is due and offer a payment link.”
Weak: “Talk to customers about their account.” (The agent has nothing concrete to do.)

First message (opening line)

The exact words the agent opens with. A strong opener is short, friendly, and says who’s calling and why. If you leave it blank, the agent will open naturally from the purpose - but writing it gives you control over that critical first sentence.

A/B openings: you can turn on a second opening line (Variant B). Calls split 50/50 between the two, so you can see which greeting gets more people to stay on the line. (Not available for Announcement agents - the fixed message overrides the opener.)

Call goal

What a successful call looks like - the outcome you want. “Book a demo,” “collect a callback time,” “confirm the address.” The agent steers the conversation toward this.

Guardrails

The hard “never do this” rules. These are absolute - the agent will not cross them.

Examples: “Never quote a specific price,” “Never promise a delivery date,” “Always confirm the spelling of the name.”
Put anything a regulator or your legal team cares about here, not in the purpose. The purpose is a goal; guardrails are the fence.

Additional info

Any extra background the agent should know - product facts, common objections and how to handle them, edge cases. Think of it as the FAQ you’d hand a new rep.

Custom variables

Placeholders like {name}, {amount}, or {due_date} that get filled in per call from your contact list. So one agent can say “Hi Priya, your payment of ₹4,000 is due Friday” to one caller and the right values to the next.

Critical rule: every variable you use here must exist as a column in your contact list. If it doesn’t, the agent will read the placeholder out loud literally (“your payment of open-brace amount close-brace”) - which sounds broken. The builder auto-detects {placeholders} you type and lists them so you can check.

Compliance pack

The rulebook the agent follows on every call - calling-time windows, Do-Not-Disturb scrubbing, how often the same person can be contacted, consent and recording notices. You pick one here and the agent obeys it automatically; the platform enforces it before a number is ever dialed. See Compliance packs. The builder suggests a pack based on your language and vertical.

Knowledge base

Attach a knowledge base - your real documents, FAQs, policies - so the agent answers from your actual content instead of guessing. If a caller asks something specific (“what’s your refund window?”), the agent looks it up and answers correctly. Optional, but it’s the difference between a confident agent and one that makes things up.

Inbound calling

Whether this agent should also answer incoming calls. (Full inbound setup lives in Inbound calls.)

Step 3 - Tools & capture

Step 2 was what the agent can say. This step is what it can do - and what it should remember from the call.

Tools

Actions the agent can take mid-conversation. Grant only what this specific call needs - every extra tool is one more thing that could fire at the wrong moment. Common ones:

Transfer to a human - hand a live, warm call to a real person. Only add this if there’s actually someone on the other end (a rota, a support desk). Without a human to receive it, the transfer dead-ends.
End call - lets the agent hang up cleanly when the conversation is genuinely done (it asks permission first). Handy to keep on.
Book a meeting / calendar - schedule directly during the call.
Send WhatsApp / SMS - trigger a follow-up message (where configured).
CRM actions & webhooks - push the outcome to your CRM, or call your own API mid-call.

Full details and setup: Tools & actions.

What to capture

The pieces of information you want pulled out of each call and saved as structured data - “was the caller interested?”, “what time do they want a callback?”, “what objection did they raise?”. After the call, these show up as clean fields you can filter and report on.

Name the keys the way your CRM does (interested, callback_at, objection) so the data lines up when you sync it out.
There’s a “Suggest fields” helper that proposes capture fields from your call purpose.

Test, then publish

Never publish blind. Two ways to try the agent first:

Simulation (test drive) - the platform role-plays a customer and runs a full call against your agent, then scores how it did. It’s fast and free, so do this first.
Test call - have the agent actually call your own phone so you hear it live.

See Test your agent for both.

When it behaves the way you want, hit Publish. Now your campaigns (outbound) and phone numbers (inbound) can point at it, and it’s live on real calls. Made a change later? Just Republish to push the update live.

Next: Prompt & flow builder → - for calls with real branching (qualify → pitch → book → handoff), where a single prompt isn’t enough.