LendingTree has built a multi-agent mortgage assistant on Amazon Bedrock to help borrowers understand options and find relevant offers through conversation. The system is designed for a regulated financial setting, where accuracy, transparent guidance, privacy, and compliance controls are required rather than optional.
The assistant uses three agents: a supervisor and two specialized workers for education and matching. The agents are coordinated with LangGraph and the Model Context Protocol, with foundation models accessed through Bedrock. The services run as containers on Amazon ECS with AWS Fargate.
The architecture reflects the complexity of mortgage decisions. Borrowers may need to compare purchase and refinance choices, fixed and adjustable rates, government-backed and conventional loans, fees, discount points, and debt-to-income rules. LendingTree wanted more than a simple FAQ bot.
The useful detail is the separation of roles. A supervised multi-agent design can keep specialized tasks clearer and make governance easier, but it still depends on guardrails, review, and reliable data sources in a high-stakes consumer finance workflow.