Harnessing Agentic Flows in GenAI at Booking.com
Hai Le | GenAI Machine Learning Manager - Booking.com
Generative AI has revolutionized how we approach travel planning, but moving from simple chatbot calls to complex, high-accuracy travel assistants requires a shift from deterministic to agentic architectures. At Booking.com, we transitioned from traditional, rigid pipelines to "Agentic Flows"—systems capable of planning, acting, and reflecting to handle sophisticated user needs like property comparisons and contextual trip support. This evolution addresses the limitations of deterministic flows, such as their lack of adaptability and inability to recover gracefully from errors.
This talk provides a practical deep dive into how we build and scale these agents, specifically by:
Developing a modular AgentHub platform that simplifies the creation of AI agents.
Implementing a robust evaluation framework that measures dialogue quality, factual accuracy, and user satisfaction at scale.
Different internal tooling and infrastructure that facilitates the end-to-end process from prototyping to productionizing an AI agent.
Attendees will gain technical insights into the architecture that democratizes AI agent development at Booking, along with key lessons on transitioning from out-of-the-box models to high-performing, in-house agentic ecosystems.
Long-Term Memory for LLM Agents: From Stateless to Stateful AI
George Chouliaras | Senior Machine Learning Scientist - Booking.com
Production LLM agents are fundamentally stateless, every conversation starts from scratch, leading to repetitive interactions, lost context, and an inability to learn from experience. We present a long-term memory framework inspired by human cognitive memory systems, organized into three types by abstraction level: episodic memories capturing compressed interaction summaries, semantic memories distilling persistent user facts and preferences across conversations, and procedural memories encoding behavioral rules extracted from successful and failed agent trajectories. For each type, we walk through its structure, generation pipeline, and key design decisions. We conclude with offline evaluation results highlighting the potential impact of memory-augmented agents.