The problem
A voice agent is easy to claim and hard to believe. Every page selling one says it is fast and sounds natural, and none of them let you hear it. The two things a buyer wants to know are whether it answers like a person and how long it leaves the caller waiting, and neither survives being written down.
What we built
A call that runs in the browser, with no account and no phone number. Speech to text, the reply and text to speech all run on Workers AI inside one Cloudflare Worker, with a Durable Object holding the call. The receptionist answers what a receptionist is asked: when they are open, whether a table is free, how large a party can be. It takes a reservation request and says a person will confirm it, because an agent that announces a confirmed booking is a complaint waiting to happen. Every turn prints its own numbers beside the transcript, taken from the voice pipeline's own metrics, on the Worker's clock.
The result
Harbour Table is a fictional restaurant and the page says so. The build is the argument: it is a receptionist you can interrupt, and the latency on the screen is the latency you just heard.