How Included Health Built Federated Healthcare Agents with LangGraph and Deep Agents

Included Health, a healthcare platform that partners with employers and health plans, built Dot, an AI-powered healthcare guide, on a federated multi-agent architecture using Deep Agents and LangGraph. The company described the system and its results in a published account.

The production architecture centers on a main LangGraph graph, called the Dot supergraph. Dot acts as the primary conversational router for transactional interactions, such as coverage questions and billing inquiries, and for navigation to care. Sub-workflows handle domain-specific member journeys including urgent care intake, appointment scheduling, finding a specialist and behavioral health.

Different product teams own different parts of the graph. Included Health added Deep Agents for consistency across services. Staff Machine Learning Engineer Rohan Bhandari said that previously, moving into a different agent made responses shorter and more brusque, without the same voice and tone. A global platform prompt for voice and tone is passed across agents, and Deep Agents' filesystem and context management let an outgoing agent summarize a conversation and pass a summary and a file path to the full history to the receiving agent.

Coverage questions are handled through a shared platform sub-agent that all Deep Agents can inherit. Clinical capabilities and services are encoded as Deep Agent skills, with a skill registry managed through a virtual filesystem using progressive disclosure.

Human handoff is a core design constraint. LangGraph's durable execution lets the agent pause when uncertain, route to a human member care advocate for a multi-turn exchange, and resume while retaining full context. Engineering Manager Kartik Darapuneni described LangGraph as the company's entire messaging platform, and said the human is helping the agent get unblocked rather than doing all the work.

Clinical oversight uses LangSmith annotation queues; every conversation goes into a queue for clinical team review, and labels are exported to the company's data warehouse. A clinical team reviews chats and confirms whether it agrees with the care spot a member was sent to, a feedback loop used to tune skill definitions and stay above a target of 95% clinical routing agreement. Multi-turn user simulation evals supported a migration from standard agents to Deep Agents across four product teams; the migration took under two weeks, with no significant regressions.

Dot launched to clients in August, which Bhandari called the smoothest launch the team had seen in years. Early metrics include a 75% lift in chat engagement, clinician agreement with Dot's care recommendations above the 95% target in graded conversations, and identification of over 99% of high-risk situations as validated by regular clinical audits.