LendingTree Develops Multi-Agent Mortgage Assistant on Amazon Bedrock

LendingTree has developed a multi-agent AI-powered mortgage assistant on Amazon Bedrock to streamline the home-buying process. The assistant employs three independent AI agents—supervisor, education, and matching—coordinated through LangGraph, the Model Context Protocol, and Amazon Nova foundation models. This setup enables the system to deliver personalized mortgage advice, answer complex questions, and match borrowers with suitable offers, all while adhering to strict financial industry regulations.

The architecture involves a consumer-facing chat interface hosted on Amazon ECS with Fargate, which communicates with the AI agents through secure, content-filtered channels. The supervisor agent orchestrates the workflow, analyzing user intent and routing tasks to specialized workers. The education worker explains mortgage concepts, while the matching worker retrieves personalized lending options by interfacing with LendingTree's internal APIs.

Operationally, the system maintains conversation context across multiple turns and agents using a PostgreSQL checkpointer, ensuring seamless user interactions. It also employs safety guardrails and policy classifiers to ensure compliance and protect user data. Since late 2025, the assistant has handled nearly 2,000 conversations, with a high rate of end-to-end resolution and significant user engagement.

The implementation has provided valuable insights into multi-agent system design, emphasizing the importance of shared foundations, reusable components, and robust safety measures. LendingTree plans to expand this architecture by leveraging shared data layers, standardized deployment processes, and agent discovery registries to scale efficiently.

Overall, this deployment demonstrates that AI agents can operate reliably in regulated industries, offering continuous, personalized guidance that enhances consumer decision-making and operational efficiency.