Agents
The AI that talks to your callers — how to create one, and what every setting in the editor actually changes.
On this page
An agent is one configured AI caller: a personality, a set of instructions, a voice, and a list of things it's allowed to do. Everything else in the product — numbers, campaigns, follow-ups, bookings — ultimately runs through an agent.
Agent → AgentsWhy you might want more than one
Most businesses start with one and never need a second. You'd add another when the job is genuinely different:
- A sales agent for new enquiries and a support agent for existing customers.
- A different agent per location or per language.
- A campaign agent with a tighter, more direct script than your inbound one.
Exactly one agent is marked default. That's the fallback used whenever nothing more specific applies — an unmapped number, a call with no routing rule. You can't delete the default agent; make another one default first.
How many agents you can have is a plan limit: 1 on Trial, 5 on Starter, 20 on Professional.
Three ways to create one
- 1
Create with AI — the fast path
A short conversation: it asks about your business one question at a time, then writes a complete agent — instructions, greeting, voice, the lot. About a minute, and the result is usually 80% right. Read it afterwards and fix the details it guessed.
- 2
From a template
A catalogue of 15 industries with several use cases each — dental clinic reception, real-estate enquiry qualification, restaurant bookings, and so on. Preview shows the opening line and the whole conversational flow before you commit.
Templates are a strong starting point, but they don't know your prices. Pair one with a knowledge base.
- 3
Blank
An empty agent. Best when you already know exactly what you want it to say, or you're copying the structure of an agent that works.
Tip
Whichever route you take, the editor is the same afterwards. Nothing is locked in by your choice of starting point.
The editor, tab by tab
Instructions
This is the agent's brain — the part worth spending your time on. Rather than one giant block of text, instructions are split into seven named sections you fill in independently:
| Section | What goes in it |
|---|---|
| Role | Who the agent is. "You are the receptionist at Lotus Dental." |
| Goal | What a successful call looks like. "Book the caller in, or take their details." |
| What to do | The steps of a normal conversation. |
| Objection handling | What to say when the caller pushes back on price, timing, trust. |
| What NOT to do | Hard limits. "Never quote a price for implants — book a consultation." |
| Rules & constraints | Policies: deposit rules, cancellation windows, who you don't serve. |
| Closing | How to wrap up and confirm. |
Each section has its own on/off switch, so you can disable one without deleting it.
Two reasons it's split up rather than one text box: you can fix one part of the agent's behaviour without rewriting everything, and short focused instructions make the agent respond faster and stay on script.
Tip
Write instructions the way you'd brief a new hire on their first day — plain sentences, concrete examples, no jargon. Avoid formatting like bullet symbols or bold markers in the text itself; the agent reads this out loud, so write it to be spoken.
If an agent was created from a template, you may see an extra Other sections block in amber. That's normal — templates use a slightly different section structure and anything that doesn't map to the seven standard names is kept intact there rather than thrown away.
Persona
The greeting — the exact first line the agent speaks when a call connects. This does more work than any other single sentence you'll write. It sets the tone, tells the caller who they've reached, and frames what the agent can help with.
A weak greeting: "Hello, how can I help you?" A strong one: "Thanks for calling Lotus Dental, this is Ava — are you calling to book an appointment or about an existing one?"
There's also a legacy free-text prompt field here. It's only used as a fallback if you've disabled every instruction section — ignore it unless you have a reason.
Voice & behaviour
Where the agent stops sounding like software:
- Voice — five voices to choose from. Listen to a couple; the difference in how callers respond is bigger than you'd expect.
- Model — which live model runs the conversation. The default is the fastest.
- Response speed — how long a pause has to be before the agent decides you've stopped talking. Too eager and it talks over people; too patient and it feels slow. If callers complain it interrupts, this is the setting.
- Start / end sensitivity — how readily it decides speech has begun or ended. Worth nudging if you get a lot of calls from noisy environments.
- Background ambience — off, quiet office, or hold music. A little background noise makes some callers more comfortable; silence can feel eerie.
- Filler phrases — short natural lines ("let me check that for you") played while a slower lookup runs, so the caller never hears dead air.
Call control
What the agent does when the conversation needs to leave it:
- Transfer to a human — a list of numbers tried in order, a plain-language condition for when to transfer ("the caller is angry, or asks for a manager, or asks about a refund over £500"), and the line the agent says before handing over.
- End-call rules — a maximum call length, a plain-language condition for ending, and a goodbye line.
- Silence handling — a three-stage escalation when the caller goes quiet ("Are you still there?" → "I'll wait a moment…" → hang up politely) instead of holding an empty line open and charging you for it.
Careful
Set the transfer condition even if you never expect it. The single worst call outcome is an agent looping on a question it can't answer while the caller gets angrier. Give it an exit.
Actions
What the agent can do, as opposed to say:
- CRM actions — schedule a follow-up, change the contact's stage, add a note, or collect specific fields from the caller (with a per-field description so the agent knows what it's asking for).
- Appointment booking — either a simple calendar (timezone, slot length, opening hours, days, how far ahead) or resource-based booking if you have multiple staff or rooms. See Appointments and Booking resources.
- External API actions — your own systems. Defined once for the workspace, then switched on per agent. See Actions.
Post-call
Extracted variables — custom fields the system pulls out of the transcript after the call ends. Give each one a name and a description ("budget: the price range the caller mentioned") and it'll appear on the call record and in the call-completed webhook.
This is how you get structured data out of unstructured conversations without listening to anything.
AI assistant
A floating chat inside the editor. Tell it what you want changed — "be firmer about taking a deposit before booking" — and it proposes a specific edit, shows you what it will change, and applies it only after you confirm. It edits individual sections surgically rather than rewriting the agent.
Testing changes
Three test surfaces, all reachable from the editor header:
- Web test — talk to it through your browser mic.
- Call test — it rings your phone.
- Chat test — type instead of talking. Fast for checking knowledge answers.
Details and limits in Testing your agent.
Things worth knowing
- Save is explicit. The editor warns you if you try to leave with unsaved changes.
- Deleting a non-default agent unassigns it from any numbers pointing at it — those numbers fall back to your default agent rather than breaking.
- The default agent cannot be deleted. Promote another one first.
- Recording is plan-gated at call time. Turning the workspace recording switch on doesn't enable recording on a plan that doesn't include it.
- CRM actions need a contact. On an inbound call from an unknown number with no matching contact, CRM writes are skipped silently rather than creating noise.