Build an AI phone agent, idea to live
A practical path, not a research project. Define the job, write the prompt, attach your number, add tools, test it, and go live. The first working version takes minutes, and every improvement after is a prompt change, not a rebuild. Flat $0.08 a minute.
- Define the jobinbound reception, booking, routing
- Write the promptpersona, rules, fallback
- Attach your numberyour carrier, one request
- Add toolscalendar, your API
- Test itcall it yourself first
- Go liveanswering real callers
The six steps, in order
Every good phone agent follows the same path. None of it is heavy, and you can have something answering on the first pass, then sharpen it from there.
Define the job
Pick one thing the agent does well: inbound reception, appointment booking, call routing, or after-hours cover. A narrow, deep agent beats a wide, shallow one, so decide what a successful call looks like before you write a word.
Write the prompt
The prompt is the agent. Set the persona, the rules it follows, and how it behaves when a caller says something unexpected. Keep it to short, natural sentences, because long answers make callers talk over it.
Attach your number
Bind one of your own carrier numbers to the agent with a single request and it answers immediately. One agent template can back many numbers, so you reuse the behaviour per customer or location.
Add tools
Connect a calendar so the agent can check availability and book in the call, and register custom functions so it can reach your own API. Keep tool results terse, just the fields the agent needs to confirm.
Test it
Call the agent yourself, or start a web call, and hear exactly what a caller would. Iterating is a prompt change, not a redeploy, so fix the rough edges in the same sitting.
Go live
Point real traffic at the number. Every call returns a transcript, summary and sentiment, so you keep refining from what real callers surface.
Start narrow, then widen
The agents that work are the ones that do one job properly. Ship a booking agent that books cleanly before you ask it to also handle billing questions and complaints. Because refining is a prompt change and one template can sit behind many numbers, widening later is cheap. Trying to do everything on day one is what makes a voice agent feel unfocused.
Go deeper on any step
Each step has its own reference when you want the detail: how numbers bind, how tools are registered, and how call data comes back.
See the AI Phone Agent API, Voice Agent API and Function calling pages, or the full docs. Coming from another platform? Nixflex is a modern alternative to Retell, Vapi and Bland. Compare them →
Developer FAQ
How long does it take to build a phone agent?
The first working version is minutes, not weeks. You write a prompt, attach a number, and the agent answers. Most of the time after that goes into refining the prompt against real calls, which is a text change rather than a rebuild, so each improvement is quick.
What makes a good phone agent prompt?
Narrow and specific beats broad and vague. Define one job clearly, say how the agent should talk, what it is allowed to do, and what to do when it is unsure. Short, natural sentences work best on a call, since long answers make callers interrupt. Start with one use case done well, then widen it.
Do I need my own phone number?
Yes, and you bring your own. Attach a Twilio or Telnyx number to the agent and it answers on it immediately. Because an agent is a template, one agent can sit behind many numbers, so you set the behaviour once and reuse it per customer or per location.
How does the agent actually book or look things up?
Through tools. Connect a calendar and the agent can check availability and book inside the call; register a custom function and it can call your own API mid-conversation. Anything that must be exact, like a time or a reference, is supplied by the engine rather than guessed, so the agent books rather than just talks about booking.
Can I test it before real callers reach it?
Yes, and you should. Call the agent yourself, or run a web call, and hear exactly what a caller would before you point real traffic at it. Since behaviour is prompt and settings, you can catch a rough edge and fix it in the same sitting.
How do I know how it is doing once it is live?
Every call returns a transcript, a summary and a sentiment score, by webhook or from GET /v1/calls, so you can watch answer quality and outcomes without listening back to each one. You keep refining the prompt from what real calls surface.
What does it cost to run?
A flat $0.08 per minute of voice, pay-as-you-go, premium model included, with your own carrier underneath. No credits and no platform fee to start; you pay for the minutes your agent is on a call.
Build your first agent today
Define the job, write a prompt, attach a number. Something answering in minutes.