The model picked a beautiful eco-lodge for the trip. That hotel does not exist, so our booking agent refused to book it. Put an LLM in charge of a booking flow and it will, sooner or later, invent a plausible property and confirm it. It reads well, the traveler pays, and the hotel was never real. That is already a 2026 bug class, and the company that deployed the agent owns every word it says (Moffatt v. Air Canada, BC Civil Resolution Tribunal, 2024, where an airline was ordered to pay after its chatbot invented a bereavement-fare policy). We built the Deterministic Travel Booking Agent so that failure can't happen by accident. The control flow is hand-written Python; the model only parses the request and words the reply. Agents advise, code decides. In the demo, one scenario asks for "Tabacon Springs Eco-Lodge," a name we fabricated on purpose. A verification gate checks every property against the (simulated) reservation system by ID before anything can be shown. Tabacon Springs isn't there, so the gate refuses to surface it. The agent doesn't improvise a confirmation. It says, plainly, that it could not confirm the property, and stops. Terminal state: abstained. 🛑 The part that outlasts model upgrades: a smarter model still hallucinates inventory under pressure. What refuses to book the thing that doesn't exist is the gate, not the prompt. If your team is weighing an agentic booking layer you can't bet the business on, we'd genuinely like to hear how you're drawing the line between what the model suggests and what the code is allowed to commit. #TravelTech #AgenticAI #AIGovernance #TMC #OTA