One-to-one service, at disruption scale.

At 10 a.m., Air Eleven cancels the 4:00 PM from Toronto to Vancouver. Five passengers now compete for two later departures: a 6:00 PM with very few seats left and a 7:40 PM with five.

Most voice-agent demos are one call with one person. Disruption recovery isn’t: every call happens at the same time, every passenger needs something different, and they all draw on the same seats. Each passenger here stands for a requirement a real airline would have.

Airy, the rebooking agent, calls all five passengers at once. Each conversation is independent; the seat inventory is not. Every call works against one authoritative reservation service.

That split is the core design rule: the model is never the source of truth for inventory. Airy can only promise what the reservation service grants through server tools. The agent decides how to talk; the service decides who gets the seat.

To test it without people on the line, each passenger is an ElevenLabs agent too, with its own persona, voice and phone number. Airy dials real numbers and talks to them over the public phone network, so the whole scenario can be rerun on demand.

Airy opens in English. When Sophie answers in French, the same live conversation adapts. ElevenLabs Scribe transcribes her French, and Airy carries on with the same voice, tools and session context, including the identity check where she spells her surname letter by letter. One workflow, now in Sophie’s language.

Priya mentions her wheelchair once. Airy verifies her and holds a seat on the 6:00 PM before transferring the live call, so the seat can’t be lost during the handoff. ElevenLabs carries the conversation context into a separate Accessibility Agent, which greets her by name and already knows her flight and her need. Priya said it once.

Marcus has a wedding tomorrow and wants cash compensation and a free business upgrade. Airy never promises the exception. It logs the request for customer care, pulls his profile from Salesforce and offers a window seat on the 6:00 PM, his saved preference. He still says yes.

The concurrency test. Maya and Daniel are on separate calls and both ask for the last seat on the 6:00 PM. The conversations are independent; the inventory isn’t. Both calls reach hold_seat, and one atomic hold in the reservation service (a single conditional SQL UPDATE) decides, so only one request can ever take the seat. Whichever hold lands first gets the seat; the other caller hears it has sold out and recovers to the 7:40 PM in the same call. In the run shown in these screenshots Maya won it. In the live take in the video, Daniel did.

Each layer owns one job. Twilio handles telephony. ElevenLabs Agents runs the conversation layer. Salesforce provides customer context. The reservation service owns inventory and atomic holds.

Inside ElevenLabs, Airy uses verification, CRM, booking and transfer tools. Trusted context is bound to each call through dynamic variables. Only after the passenger is verified do server tools like hold_seat transact against the inventory service, and each conversation record shows those tool calls succeeding.

Every prompt, tool schema and agent setting lives in the repository and is pushed to ElevenLabs by a sync script, so a prompt change is a reviewed diff with tests, not a click in a dashboard.

The result is one-to-one service at disruption scale: personalised conversations, deterministic controls, and no tradeoff between customer experience and operational integrity.

Holdline is a proof of concept, not an airline product. The airline, passengers and flights are fictional, and the inventory service stands in for a real reservation system. Taking it to production would mean Postgres instead of SQLite, batch calling with a priority engine (connection risk, fare class, loyalty), staged agent versions and real operator authentication.