A front desk that answers a public number
Our own production voice agent. It answers the phone on the work page, books real meetings on a real calendar, leaves a receipt for every call, and was audited against the catalog we sell before we sold it.
- probes in the catalog it was audited against
- severity-one findings on the first run, fixed
The problem
A voice agent demos well. Then a caller interrupts, a webhook returns a five hundred, a parameter vanishes between the console and the wire, and a retry books the slot twice. None of that shows up in a vendor test suite, because a simulated conversation mocks tool execution. We sell an audit for exactly this. The first line we pointed it at was our own.
The system
The front desk answers the number printed on the work page. It captures a lead, checks a real calendar, books a scoping call with a meeting link, and takes a message when booking is the wrong move. Every one of those is a webhook into our own runtime, and every landed call leaves a receipt before anything else happens: the transcript is written durably first, deduplicated by receipt, and only then summarized into the agent’s memory.
Two things watch it. A door watchdog probes the public route from outside our network and pages a person on the first failed probe, then again every six hours while it stays closed. A nightly reconcile counts transcripts against calls and pages when one is missing. Both exist because the failures that cost the most are silent: the dashboard green, the table empty.
The prompt, the offer card, and the tool wiring live in code and are pushed with a read-back, so a console edit cannot silently undo a fix. Eight pinned behaviors, judged on a clone with the live tool identifiers, gate every change; a sampling instrument measures single behaviors ten times over because a pass-or-fail gate cannot tell a small failure rate from a large one. The model behind the voice was chosen by that measurement, not by preference.
The first full run of the audit found six findings, 3 of them severity one: recording with no disclosure in the opening turn, retention set to unlimited, and a platform behavior that echoes credentials on a configuration read. Two were fixed the same night and verified by re-running the harness; the credential was rotated. A fourth severity-one claim did not survive our own verification pass and was downgraded rather than reported.
Receipts
- 43 probes in the catalog the desk was audited against (2026-08)
- 3 severity-one findings on the first run, fixed and verified by re-run (2026-07)
- A public phone number, on the work page, that you can call and try to break
What this means for you
If you run a voice agent, the far end of the pipe is where it fails, and nothing in the vendor’s dashboard can see it. This is the discipline the audit applies to your line and the operations retainer keeps running after we leave. The door that closed twice, and the watch that came out of it: The Far End of the Pipe. Call ours first.