Blog
Booking.com's AgentHub puts cross-sell inside the agent: how to wire ancillaries as tools
Booking.com engineers have described AgentHub, an internal platform for building modular AI agents. PhocusWire covered it on October 5, 2026. The architecture is worth reading if you build an AI travel agent, because it treats every capability, including cross-sell, as a tool the agent chooses to call.
How AgentHub builds an agent
Per Booking.com’s post, every agent shares the same five parts:
| Part | What it is |
|---|---|
| Model | The LLM that reasons and decides |
| Tools | APIs or services the agent can call to get data or take action |
| Prompt | Instructions, guidelines, and guardrails |
| Agent strategy | The execution pattern, such as ReAct or function calling |
| Output structure | A schema for the final response |
The post draws a clear line between the two layers. An agent decides which tools to use. A tool is “a deterministic software component” that performs a defined function and does not learn or adapt.
On top sit shared services: tracing with OpenTelemetry, LLM-as-judge evaluation, guardrails for prompt injection and PII, and configuration management. Booking.com reports that moving its trip-planner Q&A from a fixed flow to an agent raised answer relevance by 53% on its own evaluation set.
Cross-sell is now a funnel stage for agents
Booking.com lists three stages AgentHub supports: Discover & Plan, Search & Book, and Cross-Sell & Trip Management, with a car rental assistant and a support chatbot as examples.
The same week, Dragonpass said its AI Airport Concierge, built with Alibaba Cloud on Qwen models, matches itineraries with airport benefits, transfers, and local services, and has served more than 100,000 travellers in eight months (Dragonpass’s figure).
Two very different companies, one direction: the add-on is offered by an agent reading the trip, not by a static upsell page.
What that means for your tool layer
If the agent decides when to offer a lounge or an eSIM, the tools have to be safe for an agent to call. Booking.com’s definition is the right bar: deterministic, structured, predictable. For ancillaries that means:
- Split read and pay. A search, a quote, or a seat map is free to call and never buys anything. The paid call is a separate tool.
- Explicit yes before the paid call. The agent should treat a quote as information, not consent.
- Idempotency on every paid call. Agents retry. A retried booking must return the first result, not buy twice.
- Empty means empty. If a supplier has no lounge slot, the tool returns nothing, and the agent says so instead of guessing.
- Structured output. Price, currency, supplier, and cancellation terms in fields, so the output schema can carry them to the traveller.
We go deeper on the full flow in Build an AI travel agent that fulfills ancillaries, and on lounges specifically in Airport lounge API for AI agents.
Building it without twenty supplier contracts
Booking.com can wire internal services into AgentHub. Most teams building a travel agent cannot sign and maintain a contract per ancillary supplier.
Ancilair exposes that shelf as tools: insurance, eSIM, lounges, fast track, transfers, seats, and bags, each with a read step and an approved paid step, over MCP, REST, or CLI. Prices are quoted on the call at the supplier’s rate with 0% markup, and a contract your team already holds is used instead of ours. Start from the catalog or the Layover kit workflow.
Sources
- Booking.com: Introducing AgentHub, Booking.com platform for modular travel AI agents
- PhocusWire: Booking.com develops internal platform to build modular AI agents (Oct. 5, 2026)
- Dragonpass via ANTARA/PR Newswire: AI is shifting the membership economy (Oct. 5, 2026)
Not affiliated with Booking.com, Dragonpass, or Alibaba Cloud. Figures are the companies’ own and summarised from public posts as of October 2026.
Questions
- What is Booking.com's AgentHub?
- AgentHub is an internal Booking.com platform for building modular AI agents from configuration: a model, tools, a prompt, an agent strategy, and an output schema, with shared observability, evaluation, and guardrails. Booking.com engineers described it in a blog post covered by PhocusWire on October 5, 2026.
- Does AgentHub cover cross-selling?
- Yes. Booking.com lists Cross-Sell and Trip Management as one of the funnel stages it supports, with examples such as a car rental assistant and a customer support chatbot.
- How should an AI travel agent expose ancillaries as tools?
- Give each ancillary a read step (search, quote, or seat map) and a separate paid step that runs only after the traveller explicitly agrees, with an idempotency key. Keep tools deterministic and let the agent decide when to offer them from the trip context.