Conversational agents are easy to prototype and much harder to operate reliably in production. In this talk, Will White shares lessons from deploying LangGraph-based conversational agents at Thrive Learning, where a multi-tenant agent setup handles thousands of conversations every day.

The production system includes streaming responses, specialised agents for coaching, analytics, and chat, a retrieval-augmented generation (RAG) database, and an evaluation and monitoring layer using Arize AX. Drawing from this experience, Will shows how LangGraph can help teams move beyond fragile agent demos towards systems with clearer control flow, better observability, safer tool use, and more maintainable architecture.

The talk covers practical production patterns including state management, graph-based orchestration, monitoring and evaluations, human-in-the-loop workflows, agentic RAG, idempotent identifiers, prompt caching, summarisation, and context management. It also covers the trade-offs: where LangGraph adds real value, where it increases complexity, and how to decide whether the learning curve is worth it.

Attendees will leave with a practical checklist for designing, testing, monitoring, and evolving conversational agents in production.