Authors: Moran Beladev, Hadas Harush, Chana Ross

At Booking.com we developed AgentHub, a platform to power intelligent, modular, and multi-modal travel experiences through LLM-driven agents. Our platform provides configurable agents capable of performing a wide range of tasks, along with out-of-the-box agent management and monitoring capabilities.
With AgentHub, different teams can build and deploy AI agents in a matter of hours, based on configuration and reusable components.
Background
As AI capabilities continue to advance rapidly, agentic AI is becoming a defining force in the technology landscape. Unlike traditional generative systems, agentic AI extends large language models (LLMs) with iterative planning, autonomous decision-making, and the ability to take action through tools. By enabling direct interaction with internal systems and APIs, these agents can leverage proprietary data and orchestrate workflows end to end — unlocking deeper contextual understanding and highly personalized user experiences.

To support a wide range of GenAI use cases at Booking.com, from simple LLM-powered features to fully agentic capabilities, we built the AgentHub platform. AgentHub enables rapid agent creation through configuration-driven setups, backed by robust shared infrastructure. This approach democratizes GenAI across the company, empowering diverse roles such as UX writers, product managers, developers, and more to build and iterate quickly.
The platform supports use cases across the entire booking funnel. From the Discover & Plan stage, where travelers can chat with an AI trip planner to explore and plan itineraries, to Search & Book, enabling natural-language search, smart filtering, and automated property descriptions, and finally Cross-Sell & Trip Management, with agents like the car rental assistant and customer support chatbot. AgentHub brings intelligent, personalized assistance to every step of the journey.

Motivation and Value Proposition
As travel experiences become increasingly dynamic and personalized, our systems must evolve to match that complexity. Travelers expect seamless, intelligent interactions that understand their context, anticipate their needs, and even take actions on their behalf.
AgentHub platform was built to solve this challenge, a modular, agent-based platform designed to orchestrate both simple and complex travel workflows with flexibility and precision. It dynamically adapts to user intents, executes multi-step tasks end-to-end, and makes intelligent decisions in real-time while being able to observe and correct its outputs.
By leveraging internal and external tool integrations, real-time data, and workflow state awareness, AgentHub selects the right tools at the right time — whether it’s answering a query, recommending destinations, or booking accommodations. Agents are guided by strong guardrails and can gracefully handle execution failures.
This platform enables a shift from rigid pipelines and deterministic to adaptive, intelligent flows that reflect how real travelers think and behave — unlocking a new level of automation, scalability, and personalization in the travel journey.

Next we will deep dive into the platform components — agent templates, the Agenthub SDK we built, and platform fundamental capabilities.
Definitions
Let’s start with a few definitions to clarify the key concepts behind agentic systems:
Agent — An AI/ML-powered capability built on large language models (including multimodal models) that enables an agentic system to perceive, reason, learn, and act. Agents can autonomously carry out complex tasks by deciding which tools to use and how to use them effectively. Each agent is designed to fulfill a specific objective and may operate independently or as part of a larger workflow (e.g., a Q&A agent, a recommendation agent).
Tool — A deterministic software component, such as an API, module, or service, that performs predefined functions based on instructions from an agent. Unlike agents, tools do not learn or adapt; they provide reliable, structured operations such as querying databases, running calculations, or interacting with enterprise systems. Within an agentic workflow, tools extend an agent’s capabilities by enabling access to internal and external resources, grounding LLM outputs, and executing precise actions.
AgentHub Platform
AgentHub is the backbone of our work on domain-specialized, agentic GenAI experiences at Booking.com. It enables us to build, iterate, and operate LLM-powered agents tailored to production travel use cases at scale.
Designed with modularity at its core, AgentHub allows teams to move faster while maintaining strong control, reliability, and domain performance. The platform is built around three core components: Agent Templates, SDK to build agentic configurtions, and Platform Fundamental Capabilities.
Component #1: Agent Templates
The Agent Templates is a collection of reusable, task-oriented agent templates designed specifically for travel scenarios. These agents act as modular building blocks that can operate independently or as part of larger, multi-agent workflows.
Agents in the templates range from lightweight capabilities, such as classification or intent detection, to more advanced, autonomous agents that can reason, plan, and act using tools and memory. Each agent is purpose-built to solve a concrete problem and can dynamically decide which tools to invoke and how to use them effectively.
Agent Types & Examples
The catalog includes a growing set of predefined agent templates that teams can configure and deploy quickly, including:
- General Agent
A versatile template for conversational or structured input/output use cases. Teams can configure the prompt, model, tools, and agent strategy to build Q&A experiences, chatbots, or structured planners. This is often the best starting point for new conversational use cases. - Classification Agent
Used to detect user intent or categorize inputs in both conversational and non-conversational flows. - Summarization Agent
Designed to summarize documents or long-form content, with support for iterative summarization and handling large inputs. - Entity Extraction Agent
Extracts structured entities from free text and outputs them in a searchable format, enabling downstream search and enrichment use cases.
Component #2: Agenthub SDK
While agents differ in purpose, they all share a common structure:
- Model — the LLM responsible for reasoning and decision-making
- Tools — APIs or services the agent can invoke to retrieve data or take action
- Prompt — instructions, guidelines, and guardrails that shape behavior
- Agent Strategy — the execution pattern (e.g. ReAct, Function Calling, or custom flows)
- Output Structure — enforces a structured schema for the agent’s final response.
AgentHub provides python toolkit SDK for configuration management that allows each use case to flexibly define these components, making agents reusable, adaptable, and easy to evolve over time.

Component #3: Platform Fundamental Capabilities
AgentHub is a production-grade platform that supports the full lifecycle of GenAI applications, with built-in capabilities for quality, reliability, compliance, and control.
🔍 Observability
Every interaction is fully traceable. AgentHub logs inputs, prompts, context, LLM responses, tool calls, and agent transitions using OpenTelemetry. These logs are stored in internal systems and Arize, enabling deep debugging, monitoring, and continuous improvement.
📏 Evaluation & Quality
AgentHub includes built-in LLM-based evaluation using “LLMs as judges”. Agents are automatically and continuously evaluated across multiple dimensions, including:
- Factual accuracy
- Context relevance
- Answer relevance
- Instruction following
- User satisfaction and frustration
Evaluation results are streamed into dashboards to track performance trends and compare iterations over time. We’re expanding this further to assess full reasoning agentic trajectories and end-to-end flow quality. See here our post on GenAI Evaluation.
🛡 Guardrails
To ensure safe and compliant behavior, AgentHub integrates multiple layers of guardrails, for example, prompt injection detection, PII filtering, jailbreak prevention, use-case-specific constraints to keep responses on-topic, and more.
⚙️ Configuration Management
All agents are configurable through a unified asset management system. Teams can swap models, adjust prompts, enable tools, and fine-tune behavior without heavy engineering overhead — making experimentation and iteration fast and safe.
📊 Monitoring & Reliability
AgentHub integrates with real-time monitoring and alerting to track key system metrics such as latency, traffic, and availability. Built on a resilient, scalable architecture, the platform supports high availability, fault tolerance, and predictable performance at scale.
Our First AI Agent in Action: The Q&A Agent
One of the first AI agents we built on AgentHub was the AI Trip Planner (AITP) Q&A agent, designed to answer property-related questions from travelers in a natural, conversational way.

At its core, the Q&A agent translates free-text user questions into accurate, context-aware answers by orchestrating several intelligent steps:
- Property resolution- Identifies and maps property mentions in user input to the correct Booking.com property IDs.
- Retrieval-Augmented Generation (RAG tool)- Fetches relevant, up-to-date property information to ground responses in factual data.
- Structured property data access (tool)- Pulls detailed property attributes from internal Booking.com services.
- Geo-based reasoning (tool)- Uses geospatial tools to calculate distances and travel times — for example, proximity to landmarks, city centers, or transportation hubs.
Together, these capabilities allow the agent to reason over both unstructured user input and structured internal data, delivering precise and helpful answers.
To make this concrete, we migrated an existing deterministic Q&A flow in AITP to an agentic implementation using AgentHub. This shift led to significant quality improvements across key evaluation metrics: Answer relevance: +53%, Context relevance: +15%, Factual accuracy: +3.3%, LLM Rejection rate: –64%.
These gains highlight the value of agentic decision-making: allowing the system to choose the right tools, retrieve the right context, and reason holistically about the user’s question.
Key Takeaways & What’s Next
AgentHub is still evolving, and we see this as just the beginning. Looking ahead, we’re focused on expanding the platform in several directions:
- Advanced multi-agent collaboration, where specialized agents work together on complex tasks.
- Multimodal capabilities, incorporating images and other non-text inputs as platform capabilities.
- Long-term memory and personalization, allowing agents to retain context across interactions and deliver more coherent, tailored experiences.
- Enhanced tool-calling methods and techniques, enabling agents to select, use, and coordinate tools more effectively and reliably.
- Agent-harness optimization, treating the complete set of configuration parameters as a search space and systematically optimizing them for quality, efficiency, and cost.
Our long-term vision is to power a new generation of intelligent, context-aware travel agents that continuously learn, adapt, and elevate the traveler journey across every touchpoint, from inspiration to booking to post-trip support.
Introducing AgentHub: Booking.com Platform for Modular Travel AI Agents was originally published in Booking.com ML & DS Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.





