Vibe coding a voice agent: describe it, test it, ship it

"Vibe coding" started as a way to describe building software by talking to an AI: you say what you want, look at what comes back, and keep steering until it's right, without writing the code yourself. It works especially well for voice agents, because a voice agent is mostly a description of a job: who it is, what it should do, what it must never do, and what to do when it's stuck.

Apollo is built around that idea. You describe the phone agent you want, and Apollo builds the voice, the call flow and the backend behind it. This post walks through what that looks like in practice.

Why voice agents used to be hard

Getting a phone agent to do real work used to mean stitching together a lot of parts:

• A phone number and a way to route calls to software
• Speech recognition, a language model and a voice, each with its own account and keys
• A prompt that keeps the agent on task, plus a flow chart for anything structured
• A server for the things the agent should do, like checking availability or saving a lead
• Somewhere to store what callers said, and a way to tell your other systems

The conversation itself was the easy part. Everything around it is what took weeks. Vibe coding a voice agent means all of that comes from the same conversation you'd have briefing a new hire.

1. Describe it

Start with what you'd tell a new receptionist on day one. Be concrete about the job, not the technology:

"A receptionist for my dental office. Book cleanings into my calendar, save every new patient to a table, and text me when someone cancels."

Apollo asks for what it needs to do the job well: your hours, your prices, where bookings should go. From one message like that, it builds:

• The voice agent: its role, greeting, tone, and the rules it follows, with your hours and policies.
• Actions: things it does mid-call, like checking a calendar. Each one is hosted code Apollo drafts for you, running in a sandbox, that can call any HTTPS API with secrets you save.
• Data tables: spreadsheet-style tables the agent reads and writes during calls, like a Patients table filled in after every call.
• Webhooks: signed events to your own systems when calls end, records arrive or rows change.
• Readiness tests: simulated callers it has to handle before it goes live.

If you'd rather not start from a blank page, pick an industry starter instead: a clinic front desk, a salon receptionist, a home-services dispatcher, a restaurant host and more. Each comes with its role, call flow and safety rules, and you only type your business name. See use cases for what each one does.

Tips for a good description

• Say what success looks like. "A booked appointment with name, number and preferred time" beats "help callers".
• Say what it must never do. "Never quote a firm price" or "never confirm a time yourself".
• Say when a person takes over. "Billing questions go to a person."
• Give it the facts. Hours, prices, service area, policies. An agent only knows what you tell it, and a good one says it doesn't know rather than guessing.

2. Test it

As soon as the agent exists, you can call it in your browser. No number needed, and test calls don't use minutes. Talk to it like a real caller would: interrupt it, change your mind, ask something it shouldn't answer.

Then run the readiness tests. An AI caller plays the hard cases (insisting on a person, going off topic, probing for the agent's instructions, plus the cases specific to your industry) and each conversation is graded. A scenario only passes if it passes all 3 runs, so one lucky answer isn't enough. There's a full checklist in how to test a voice agent before it takes real calls.

Change it by chatting

When something's off, say so the same way you described it:

• "It reads prices too fast. Slow down when you say numbers."
• "If someone asks about whitening, mention the new-patient offer."
• "Ask for the date of birth before the reason for the visit."

Every save is live on the next call. If you prefer to work directly, you can edit the prompt in plain words, or open the call flow and redraw the steps: what to say, when to branch, which tool to call and when to hand off.

3. Ship it

When it passes, put it on a phone line. On a paid plan, search an area code and get a number in a minute, or keep the number your customers know and forward it to the agent. Here's how forwarding works.

From then on every call is logged with its recording, a transcript and a short summary, and the details the agent captured land in your Inbox as structured records you define, like names, reasons and appointment times. Dashboards chart calls, outcomes and captured data, per agent or across the workspace.

When you need more than the chat

Vibe coding gets you most of the way, and the rest is still there when you want it:

• Actions are real code. Read and edit what Apollo drafted. An agent can use up to 10 actions, depending on your plan.
• The call flow is editable. For structured calls like intake, draw the exact steps.
• There's a public API. Everything the chat can do is a documented API, so your own tools can build and manage agents.
• Webhooks are signed. Verify each delivery before you trust it, and retries are handled for you.

What vibe coding doesn't replace

It doesn't replace knowing your business. The agent is only as good as the facts and rules you give it, and the quickest way to improve one is to read real transcripts and fix what you see. It also doesn't replace a person for the calls that need one, which is why every agent can transfer a live call to your team.

Try it

1. 1Describe the agent you want, or pick a starter.
2. 2Call it in your browser and run the readiness tests.
3. 3Keep chatting until it's right.
4. 4Get a number or forward your line.

Building and testing are free. Start building free, see pricing for when you're ready for a number, or browse the FAQ.