WED3:45 PM
A true story, told in timestamps
A two-engineer health-diagnostics studio asked us at 3:45 in the afternoon whether they could ship a production-grade phone agent the same day.
This is what happened next, receipt by receipt.
Names, numbers, and identifying details changed. Everything else is verbatim.
Chapter 1 · Wednesday 3:45 PM
He was mid-trial on an AI support product priced at $0.99 per resolved outcome, and he'd already run the numbers every technical founder runs:
What he wanted was something at a tenth of the cost that needed an hour a week of his attention. We told him honestly: at ten calls a day, the incumbent was a fine place to start.
We recommended the competitor.
He switched anyway.
Chapter 2 · Wednesday 7:20 PM
He didn't sit down and build a voice agent. He pointed Claude Code at the VoiceRun CLI and docs, gave it a goal, and went back to his other work.
The platform is meant for agents,
not us mere humans.
— Derek, later that night
Chapter 3 · Wednesday 8:21 PM
“hi there… actually never mind, thanks.”
First session ever · a 12-second mic check · 8:21 PM
Seventeen minutes later came the first full simulated call — and the agent's first skill wasn't succeeding. It was refusing to fail badly: a caller it couldn't find, a date-only search declined three times, a clean handoff instead of a hallucinated booking.
The founders traded model recommendations over text, and minutes later the release log shows the voice settled on their pick, Cartesia. At 8:52 PM, this happened:
Oh, hi. I need to move my appointment tomorrow.
What phone number or email was the booking made under?
Yeah. It's six five zero, five five five, oh one four two.
I found one: your appointment tomorrow, Thursday at one o'clock with Jordan Avery. Is that the one you want to move?
Could we do sometime next week? Mornings are better.
Done — you're moved to Friday the thirty-first at ten in the morning, and you'll get a confirmation email.
Chapter 4 · Wednesday 9:18 PM
Within half an hour he'd booked a real appointment with himself, dropped the agent's phone number in the thread, and dared the room: “Book an appointment w me then call this number to reschedule it 😎”
Can I reschedule it again? Actually, I— I actually can't do Saturday.
Sure — what phone number or email is the booking under?
Yeah. Do you have anything Monday?
You're all set for Monday the third at one in the afternoon with Jordan Avery.
What's in a Big Mac?
I can't help with that here, but the team will follow up by email today.
A real phone-line reschedule in 94 seconds. A cancellation in 59.
A Big Mac question, politely declined.
Chapter 5 · Thursday 1:28 AM
While the studio slept, the release log kept moving. A new simulated caller for prospective-client FAQs went up at 1:28 AM, and the small hours filled with scope probes — late-arrival policies, pricing questions, things outside the agent's mandate. Nobody was on the phones. The tests were.
Oh, hey. Um, I'm thinking about booking an appointment, but I have a few questions first. So how late can I actually be before you guys cancel my appointment?
Under five minutes, we'll almost always still get you in. Past seven minutes, we do have to cancel it, so if you think you'll be late, calling ahead really helps.
The agent spent the night
taking its own exams.
Chapter 6 · Thursday 11:55 AM
At 11:55 AM a live caller said they were running late — and got a polite but generic handoff. That was the rough edge. Within the hour, the release log shows a first-class running-late flow: the agent looks up the booking and messages the technician directly on Slack:
Heads up — your 1:00 PM client called: they're running about 5 minutes late. Told them under five is usually fine.
So do they actually know? Did you really send something, or are you just telling me that?
Yes — the message was delivered to Jordan Avery on Slack. Anything else about your appointment?
That answer is honest by construction: fifteen minutes behind the feature, in the same burst of releases, shipped a deterministic evaluator that cross-checks every claimed notification against the real send log. Claims without receipts fail the eval.
Chapter 7 · The receipts
Every number below is read straight from the platform — releases, sessions, transcripts, evaluator runs — timestamp for timestamp against the text thread.
“We definitely want to use it.”
— the Customer's co-founding engineer · Thursday 10:28 AM
A true customer story from July 2026, reconstructed from the actual text thread and the platform's own release, session, and evaluation records. Names, contact details, and identifying references have been changed or redacted. voicerun.com