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

Buying a house is confusing enough without lenders throwing around terms like debt-to-income ratio and origination fees. LendingTree, which has spent 25 years matching borrowers with lenders, decided the fix wasn't another basic chatbot. It built a three-agent system on Amazon Bedrock: a Supervisor agent that reads intent and routes traffic, an Education worker that explains mortgage concepts using retrieval-augmented knowledge bases, and a Matching worker that pulls real rate and eligibility data from LendingTree's own APIs.

The architecture is deliberately modular. Each agent runs as its own containerized service on Amazon ECS with Fargate, talking to each other over the Model Context Protocol so any one piece can be updated or rolled back without touching the others. The Supervisor, built with LangGraph as a state machine, doesn't just guess what to do — it produces an explicit execution plan and picks models based on the job, using the heavier Nova Pro for complex reasoning and the cheaper Nova Lite for routine chatter. That's a cost-control trick as much as a performance one.

Because this touches regulated financial advice, LendingTree leaned hard on Bedrock's built-in guardrails: PII redaction, prompt-injection screening, and a parallel LLM-based safety classifier that runs alongside the guardrails rather than after them, so compliance checks don't slow down responses. Conversation state persists in PostgreSQL via a LangGraph checkpointer, meaning someone can start a chat, close the browser, and pick the thread back up days later without repeating themselves.

The numbers from launch through Q1 2026 tell a fairly striking adoption story. Nearly 1,960 conversations, over 12,100 messages, and — more tellingly — engaged users sticking around for 10-plus messages across 9 minutes. Early on, three-quarters of chats were purely educational, people asking what an FHA loan even is. As usage matured, over half of recent conversations shifted toward rate comparisons and prequalification, the stuff that actually converts. Human escalation happened in only about 3% of cases, a containment rate LendingTree treats as proof the system works as a standalone advisor, not just a filter in front of a call center.

The harder lessons were less about infrastructure and more about agent design. Semantic chunking of documents beat fixed-size chunks for retrieval accuracy. When two knowledge bases disagreed, LendingTree had to build source-prioritization rules so internal product data overrides generic mortgage content. And workers needed full conversation history passed through MCP requests, not just the latest message, or their answers felt disconnected from what the user had actually asked three turns earlier.