Expedia Group Lays Out How It Governs AI Agents: Multi-Agent Trips, "Evals Are the New PRD" and Release Tollgates

Expedia Group has published the internal operating model it uses to build AI agents. It is a useful signal of how the second-largest OTA expects agentic features to reach travelers and partners, and of the guardrails it has set around them.

The post, by Xavier Amatriain on the Expedia Group Technology blog, says agents are already appearing across the business: traveler-facing trip planners, partner tools, internal copilots and log readers.

The product thesis: many agents, not one

Expedia Group says it does not expect one monolithic agent to do everything. Instead, it describes "a composed, multi-agent system". Specialised agents handle discrete jobs such as understanding intent, discovery, comparing properties, building itineraries, answering booking questions and helping when plans change. They are then orchestrated into a continuous "whole-trip experience".

The company says narrow sub-agents are easier to test, secure and improve.

Two principles stand out for anyone selling through Expedia:

• Travelers keep the final click. Agents "can inform, compare, recommend, and help a traveler decide", but the traveler should retain the final action on consequential decisions such as completing a booking.
• Proprietary assets are the differentiator. The company lists trusted supply, rich and current content, marketplace signals, first-party traveler understanding and accountable servicing as what separates its agents from "generic AI".

"Evals are the new PRD"

Expedia Group argues that evaluations should replace the product requirements document as the way teams specify behaviour. Before building, teams define expected outcomes, unacceptable failure modes, security and privacy requirements, and risk-based release criteria. Production signals then feed back into the evaluation suite.

The four-layer operating model

1. Communication and culture: AI Champion groups embedded in teams, an Agentic Leads Guild, and shared hubs so teams reuse patterns rather than rebuild them.
2. Platform and tooling: a common GenAI and agentic "fabric" with governed model access, shared evaluation tooling, and Model Context Protocol (MCP) governance. It is built for model optionality, so tasks can be routed on quality, latency, cost and risk.
3. Governance: every AI initiative that goes live, including agents, must be registered and risk-rated. A lightweight internal helper and a traveler-facing trip-planning agent follow different paths.
4. Enforcement: risk-tiered Agent Release Tollgates that shift evaluation, security, cost and observability checks earlier in development. They are already tracked through Jira automation, with plans to automate execution.

Measuring impact

Agents are judged on task completion, reduced manual effort or cycle time, adoption, cost and latency, and whether they move a real traveler or business metric. The post says a clever agent that nobody uses, or that costs more than the work it replaces, is "a useful experiment, not a template to scale."

Why it matters

For hotels and partners, the headline is that Expedia's agentic future is built on supply and content quality, with orchestrated specialist agents deciding what gets surfaced. The same week, Expedia's head of organic and agentic search described a dedicated business-to-agent team. Together they suggest Expedia is working both sides: its own agents, and how external agents see its inventory.