How LendingTree built a multi-agent mortgage assistant on Amazon Bedrock

LendingTree built a production multi-agent mortgage assistant on Amazon Bedrock so customers can get personalized mortgage guidance around the clock. The system coordinates three agents rather than routing every request through a single model call, and it is designed to operate under the compliance constraints that apply to financial services.

The stack combines LangGraph for multi-agent orchestration, the Model Context Protocol for tool and context wiring, and Amazon Nova models hosted on Amazon Bedrock. Guardrails are built into the Bedrock path so policy and safety checks sit next to generation instead of as a separate afterthought. The three agents share work so mortgage questions can be broken into specialized steps while still returning a coherent answer to the borrower.

For engineers, the interesting piece is not a demo chatbot but a production pattern: agent roles, shared state via LangGraph, and MCP-based context under a managed model layer. That layout makes it easier to reason about which agent owns which task, where tools attach, and where guardrails must fire before money-adjacent advice leaves the system.

In market terms, mortgage shopping is high-intent, regulated, and often slow outside business hours. A 24/7 multi-agent assistant on Amazon Bedrock puts LendingTree in the same lane as other financial firms using managed generative AI for customer guidance, while competing on always-on personalization without standing up its own model-hosting stack.

What to watch next is how the three-agent split holds up under real traffic: latency across LangGraph handoffs, MCP tool reliability, and whether Bedrock guardrails plus Amazon Nova stay within compliance review without starving answer quality. Builders in fintech should treat this as a concrete reference for multi-agent design on Bedrock rather than a greenfield research architecture.