Skip to content

Agent Studio

Build her. Test her. Publish her. Roll her back.

Your AI agent should be a product you own, not a black box somebody else operates on your behalf. Every vendor asks you to trust theirs. This one hands you the controls.

Built, not yet deployed

Agent Studio sits behind the same gating as the AI voice agent and is not part of the shipping product. Everything described here exists and is being hardened — it is not a roadmap sketch — but you cannot buy it this quarter, and we are choosing design partners rather than selling it.

The dialer, compliance controls, IVR, recording, reporting and messaging described elsewhere on this site are shipping today.

The usual arrangement

  • The vendor writes your agent's script.
  • Changing a price means raising a ticket.
  • You find out it went wrong from a customer.
  • Nobody can tell you why it said that.
  • The bill arrives after the calls do.

What you get instead

  • You write her, in your words, in your console.
  • Change a price and publish a new version.
  • Hear her yourself before any customer does.
  • Replay any call with every decision she made.
  • Set the daily ceiling before the calls start.

Build

Give her a character and a brief

Persona

Name, voice, gender, tone and speaking pace. Formal, warm or assertive. A receptionist does not sound like a collections agent, and neither should sound like a demo.

Knowledge

Your catalogue, prices, opening hours, promotions and policies — in your words, scoped to your tenant. Facts are read from here, never invented by the model.

Conversation design

Greeting and permission, discovery, qualification, presentation, objection handling and close, each with its own required statements and permitted actions.

Language profile

Which language, which dialect, which register. A profile rather than a translation, so the agent is written in the language rather than into it.

Prove

Hear her before your customers do

The gap between "the transcript looks fine" and "she sounds right on a phone line" is where AI voice projects fail. So you listen first, and afterwards you can listen to anything that already happened.

Test-call simulator

Hear her before your customers do. Run the flow end to end and listen to what she actually says, not a transcript of what she was meant to say.

Conversation replay

Play back any call with the full transcript alongside every decision the runtime made — which state she was in, which tools she was permitted, what she chose.

Guardrail interventions

See exactly where she was stopped and why. A blocked utterance is a logged event you can read, not a silent correction.

Language certification

Each language profile is assessed on its own — dialect authenticity, pronunciation, register, code-switching — rather than assumed to work because English does.

Control

Every switch is on your side of the glass

Versioned publishes

Every publish is a numbered version and every version is kept. If today's change made her worse, roll back to yesterday.

Kill switch

One button, immediate stop. Not a support ticket, not a business-hours escalation — a control your supervisor holds.

Budget ceiling

Set the daily spend cap. When it is reached she stops. You will never open an invoice you did not expect.

Capacity limits

Concurrent calls, calls per hour, calls per day and maximum call length, so an AI campaign can never outrun what your operation can absorb.

No usage surprises. You set the daily budget. When it is reached, she stops. You will never open an invoice you did not expect — which is not a promise most AI voice vendors are willing to put in writing.

For BPOs and resellers

One studio, many clients

If you run other people's brands, an agent you cannot configure is an agent you cannot sell. Each client is a separate tenant with its own agents, its own knowledge, its own voice and its own budget ceiling.

Separate agents per client

Nothing is shared between tenants — not knowledge, not personas, not call history.

Your brand on the console

White-label the platform so your clients see you, not us.

Per-client ceilings

Budget and capacity limits are set per tenant, so one client cannot spend another's headroom.

Questions

Owning the agent

Who configures the agent — us or you?
You do. Agent Studio is a workspace in the console: persona, knowledge, conversation design and language profile are yours to set, test and publish. That is the difference between a platform you operate and a vendor you file requests with, and it is the whole reason the studio exists as a product surface rather than an internal tool.
What happens if a change makes the agent worse?
You roll it back. Every publish is a numbered version and every version is retained, so reverting is a click rather than an attempt to remember what you changed. Versions and releases are separate records, which means you can promote a version deliberately rather than having the latest edit go live the moment you save it.
Can we hear the agent before customers do?
Yes, and you should. The test-call simulator runs the flow end to end so you hear what she actually says. After that, conversation replay lets you play back any real call with the full transcript and every decision the runtime made alongside it — which state she was in, which tools she was allowed, and what she chose.
How do we know when a guardrail stopped her?
Guardrail interventions are logged as events you can read, with the utterance that was blocked and the rule that blocked it. A policy that silently corrects the agent teaches you nothing; one that shows its working lets you fix the underlying script.
How do we avoid an unexpected bill?
You set a daily budget cap and capacity limits — concurrent calls, calls per hour, calls per day, maximum call length. When the budget is reached, she stops. The ceiling is a setting you control rather than an alert that arrives after the money is spent.
Is Agent Studio available today?
No. It is built but not deployed, behind the same gating as the AI voice agent itself, and excluded from the shipping product. We are selecting design partners rather than selling it. Everything on this page describes what exists and is being hardened — not a roadmap sketch, and not something you can buy this quarter.

Want to shape it?

Design partners get the studio built around how they actually work, rather than being handed something designed for somebody else. Tell us what you run.