Agent library

Thirty-eight agents, ready to deploy.

Every agent below has already run real calls for someone. Clone one into your workspace, swap in your price list, your policy and your voice, rehearse it — and it is yours. No blank prompt, no six-week discovery phase.

38 agents, 6 families 42 languages, 9 accents One clone away from production
Anatomy

An agent is five things, versioned together.

An agent is not a prompt and a hope. It is a prompt, a set of tools, a guardrail policy, a voice and a memory scope — bound as one artefact so the thing you approve is the thing that runs.

01 / PROMPT

Instructions & persona

The system prompt, tone rules, language behaviour and the fallback phrasing. Written in the Studio, diffable against the last version, reviewed like any other change.

02 / TOOLS

What it can actually do

Typed tool calls — look up a price, check a calendar, start a payment, open a ticket. Each one declares its inputs, its failure mode and what the agent should say when it breaks.

03 / GUARDRAILS

What it must never do

Hard limits mapped to your policy clauses: no disclosure to third parties, no clinical advice, no quote below list price. A guardrail is a stop, not a suggestion.

04 / VOICE

How it sounds

Voice, pace, pause length and the barge-in behaviour that decides whether the agent feels human or feels like a phone tree. Clone a preset or bring your own recorded voice.

05 / MEMORY

What it remembers

Caller history, open tickets, order state and prior objections, scoped per workspace and per region. Memory expires on your schedule, not ours.

Ask us

Not sure what to clone?

Send us three recordings. We will tell you which agent fits, and which one you would be wrong to pick.

The library

38 agents across six families. Filter, then clone.

Containment rates are the median reported across workspaces running that agent in production for at least 90 days. Your number will differ — the point of the clone is that you find out fast.

Qualify-Buyer
Sales & Dev · v3.4

First-touch discovery for high-ticket sales: budget, authority, need and timeline, then hands the rep a scored lead with the sentences that produced the score.

EN · HI · TA · AR · ES

containment 71%Clone & adapt →
Site-Visit-Booker
Sales & Dev · v2.8

Real-estate and showroom visits: qualifies the project, reads live rep availability, books the slot and texts directions with a map pin.

EN · HI · AR

containment 76%Clone & adapt →
Renewal-Nudge
Sales & Dev · v4.1

Subscription and policy renewals with usage context, save offers kept inside margin rules, and an opt-out that is honoured on the first ask.

EN · HI · ES · PT

containment 63%Clone & adapt →
Demo-Setter
Sales & Dev · v2.2

Books product demos for SDR teams, screening for team size and evaluation timeline before spending a calendar slot on the wrong fit.

EN · DE · FR

containment 69%Clone & adapt →
Trial-Activator
Sales & Dev · v1.9

Calls trial signups who stalled at setup, walks them past the first real blocker on the phone, and books the onboarding session.

EN · ES

containment 58%Clone & adapt →
Quote-Chaser
Sales & Dev · v3.0

Follows up sent quotations, absorbs pricing objections without inventing discounts, and records the real reason the deal stalled.

EN · HI

containment 61%Clone & adapt →
Upsell-Scan
Sales & Dev · v2.5

Reviews account usage before the call and offers the next tier only where the evidence supports it — no upgrade pitch on an account that is already over-served.

EN · PT

containment 66%Clone & adapt →
Triage-Line
Support · v5.2

The front door: identifies the issue, resolves the simple 40% outright, and routes everything else with full context attached to the ticket.

14 languages

containment 82%Clone & adapt →
Order-Status
Support · v3.7

Where-is-my-order calls answered with live carrier state, revised ETAs and a proactive call when a delivery is going to slip.

EN · HI · ID · TH · VI

containment 88%Clone & adapt →
Refund-Guide
Support · v2.9

Explains entitlement clearly, starts the refund where it is owed, and sets an honest timeline instead of the one the customer wanted to hear.

EN · ES

containment 74%Clone & adapt →
How-To-Helper
Support · v3.1

First-line product help drawn from your documentation and your own resolved tickets, so the answer matches the product you actually shipped.

EN · DE

containment 77%Clone & adapt →
Warranty-Claim
Support · v2.4

Registers claims, captures serial numbers and photos by SMS link, and books the service slot in the same call.

EN · HI · AR

containment 68%Clone & adapt →
Password-Rescue
Support · v1.6

Identity-safe access recovery that never reads a credential aloud and never asks for one it does not need.

EN

containment 85%Clone & adapt →
Churn-Save
Support · v3.8

Retention calls with churn-signal context, save offers inside margin rules, and an honest exit path when the customer has already decided.

EN · ES · PT

containment 57%Clone & adapt →
Dunning-Soft
Collections · v4.4

Early-bucket reminders that protect the relationship: no threats, no invented penalties, always a specific date to work with.

EN · HI · TA · AR

containment 72%Clone & adapt →
Promise-Capture
Collections · v2.7

Secures a promise-to-pay with amount, date and channel, then schedules the check and follows up if the commitment slips.

EN · HI

containment 79%Clone & adapt →
KYC-Verify
Collections · v3.3

Identity and income verification with per-jurisdiction disclosure rules baked in, so a compliance officer can read the rule and the transcript side by side.

EN · HI

containment 81%Clone & adapt →
Restructure-Offer
Collections · v2.1

Presents EMI restructure options inside pre-approved policy bands, and escalates rather than improvising when the ask falls outside them.

EN

containment 64%Clone & adapt →
Field-Visit-Confirm
Collections · v1.8

Confirms doorstep visits so collectors do not spend an afternoon arriving at empty houses, and reschedules the ones that slip.

EN · HI

containment 83%Clone & adapt →
Hardship-Route
Collections · v3.0

Detects distress signals — bereavement, medical bills, job loss — and routes to the human hardship desk before any collection pressure is applied.

EN

containment 61%Clone & adapt →
Triage-Clinic
Healthcare · v4.0

Symptom intake to the right department with red-flag escalation in a single turn. It never diagnoses and never names a medication.

EN · HI · AR

containment 73%Clone & adapt →
No-Show-Prevent
Healthcare · v3.6

Confirmation and reschedule calls two days out, with prep instructions and transport guidance for early appointments.

EN · HI · TA

containment 84%Clone & adapt →
Refill-Reminder
Healthcare · v2.3

Prescription refill reminders and pharmacy routing, with every clinical question handed to a human before it is answered.

EN

containment 78%Clone & adapt →
Insurance-Verify
Healthcare · v2.8

Eligibility and coverage checks before a diagnostic slot is booked, so the front desk is not the one delivering bad news.

EN

containment 75%Clone & adapt →
Lab-Prep
Healthcare · v1.7

Fasting windows, medication holds and arrival times, delivered in the patient's language and repeated back for confirmation.

EN · HI

containment 86%Clone & adapt →
Post-Op-Check
Healthcare · v2.0

Structured follow-up after discharge, escalating warning signs to the on-call clinician within the same turn.

EN

containment 62%Clone & adapt →
COD-Confirm
Logistics · v3.9

Confirms cash-on-delivery orders before dispatch, repairs addresses and converts to prepaid where the numbers justify the offer.

EN · HI · TA · BN

containment 80%Clone & adapt →
Route-Dispatch
Logistics · v2.6

Calls drivers and hub supervisors with route changes, takes an explicit acknowledgement and logs every exception.

EN · HI

containment 77%Clone & adapt →
Delivery-Reschedule
Logistics · v3.2

Moves a delivery slot without a callback, updates the carrier and confirms the new window to the customer in the same call.

EN · HI · ID

containment 85%Clone & adapt →
Address-Repair
Logistics · v2.4

Rescues undeliverable addresses with landmark, floor and pincode verification, then pushes the correction to the carrier.

EN · HI

containment 79%Clone & adapt →
Fleet-Alert
Logistics · v1.5

Driver-safety and maintenance alerts delivered by voice, because the SMS about a brake fault is the one nobody reads.

EN

containment 71%Clone & adapt →
POD-Verify
Logistics · v2.2

Confirms proof-of-delivery with the actual receiver and closes the order cleanly, instead of arguing about it three days later.

EN · HI

containment 82%Clone & adapt →
Screen-Screen
Recruiting · v3.5

First-round screening against a role scorecard, scheduling only the candidates who clear the bar you wrote down.

EN · HI

containment 68%Clone & adapt →
Interview-Booker
Recruiting · v2.9

Coordinates panel availability across time zones and confirms with both sides, rescheduling without a human in the middle.

EN

containment 81%Clone & adapt →
Offer-Followup
Recruiting · v2.2

Checks in after an offer, answers logistics questions, and flags hesitation to the recruiter the same day rather than the week after.

EN

containment 70%Clone & adapt →
Reference-Check
Recruiting · v1.9

Structured reference calls with consistent questions and verbatim capture, so references stop being a formality nobody documents.

EN

containment 74%Clone & adapt →
Bulk-Shortlist
Recruiting · v3.1

High-volume drives: screens thousands against the rubric, ranks them consistently, and books only the shortlist.

EN · HI

containment 66%Clone & adapt →
Onboard-Nudge
Recruiting · v2.0

Reminds new hires about documents, day-one logistics and equipment — before they go quiet and disappear into the gap.

EN

containment 83%Clone & adapt →
Build your own

Nothing here fits? Start from a blank page in an hour.

The clone button is a shortcut, not a cage. Every library agent is built from the same public primitives you get on day one — so a bespoke agent is a difference in configuration, not a different product.

Describe the job

Write the outcome you want — “confirm the order and fix the address” — and the Studio proposes a prompt, a tool set and a guardrail baseline you can edit line by line.

Wire the tools

Attach integrations from the directory, or define your own HTTP tool with a schema and a failure script. Nothing is callable until you approve it.

Choose a voice

Pick a preset, adjust pace and pause, then listen to the agent read your own worst-case script back before you ever dial a real number.

Rehearse, then publish

Run the synthetic crowd, read the failure list, fix, and sign the version. The signature is what goes live — not the draft.

Versioning

Agents are versioned like code, because they are production systems.

Nothing changes on a live number without a version, a diff and a reviewer. If a change makes things worse at 2pm on a Tuesday, you roll it back in one click and know exactly which sentence did the damage.

BRANCH

Change on a branch

Every edit happens on a branch of the agent, not on the live instance. You can A/B a branch against 5% of live traffic before you commit to it.

DIFF

See the sentence, not the file

Diffs are semantic: the prompt change, the new tool, the moved guardrail, the voice tweak. A reviewer reads the change, not a text file.

SIGN

Two people, one signature

Publishing requires a rehearsal run that passed and a named approver. The signature, the run id and the approver are all stored on the version.

ROLL BACK

One click, no ceremony

Roll back to any signed version. Live traffic moves within seconds and the incident timeline records who approved what and when.

FREEZE

Pin a version to a number

Different numbers can run different versions. Pilot one, keep the rest on the frozen version your risk team already approved.

SUNSET

Retire with a paper trail

Deprecated agents keep their audit history, their rehearsal reports and their transcript archive for as long as your retention policy demands.

The versioning rule. The live number only ever runs a signed version. A signed version only ever comes from a passing rehearsal. A draft can never be dialled — not by an admin, not by support, not by us.
Guardrails

Every agent ships with limits it cannot talk its way out of.

Guardrails are enforced in the runtime, not in the prompt. A model that has a bad generation cannot override a rule that was never given to it.

0
Adversarial personas
Used on every publish, not just launch
0
Unreviewed publishes
A version without a signature cannot dial
0
Longest escalation handoff
From caller request to a human voice
0
Consent captured
Recording notice spoken before any data

Hard stops, not heuristics

Disclosure to third parties, clinical advice, price below the floor, out-of-window dialling — these are refusals enforced before generation, and they cannot be renegotiated by a persuasive caller.

Escalation that actually lands

When the guardrail says “hand over”, the summary is already written, the live call is warm-transferred, and the human's screen carries the transcript so far.

Region-aware by default

Consent wording, recording notices and retention windows follow the caller's jurisdiction, not the workspace default. You configure once; the runtime adapts per call.

Start

Clone one agent. Put it on one real number.

Tell us the job and we will point you at the closest agent in the library. You will have it rehearsed and live on a test number inside a week.

Library agents are included on every plan. Cloning is free; only connected calls are billed.