Included Health Deploys AI Care Guide Using LangGraph

Included Health deployed its new conversational assistant, Dot, using a federated multi-agent architecture powered by LangGraph and Deep Agents. The system, which launched to clients in August, replaces rigid decision-tree routing with a flexible supergraph capable of handling complex medical, financial, and administrative queries. By utilizing LangGraph, the company allows different product teams to independently build and manage specific sub-workflows, such as urgent care intake, appointment scheduling, and specialist matching.

To maintain a consistent voice and tone across these disparate services, the engineering team integrated Deep Agents. This framework utilizes a shared filesystem and built-in context management, allowing outgoing agents to pass conversation summaries and full history file paths to receiving agents. This setup prevents users from having to repeat themselves. Furthermore, the architecture leverages progressive disclosure; Dot initially receives short descriptions of available skills from a registry and dynamically loads full skill files only when relevant, optimizing model efficiency.

A core design constraint of the system is its human-in-the-loop integration. Using LangGraph's durable execution, the system can indefinitely pause its graph when encountering uncertainty, hand the conversation to a human care advocate, and resume later with the full context intact. For clinical safety and oversight, Included Health routes conversations to LangSmith annotation queues. A clinical team reviews these chats to verify routing accuracy, exporting labels to a data warehouse to generate operational metrics.

The deployment has yielded significant performance gains. Included Health reported a 75% lift in chat engagement. Clinicians agreed with Dot's care recommendations well above their 95% target, and the system successfully identified over 99% of high-risk situations. For AI practitioners, this implementation demonstrates how a federated agent architecture can solve the trade-off between personalization and scale in highly regulated, high-context industries. Using multi-turn user simulation evaluations, the team even completed their entire Deep Agents migration in under two weeks without regressions.