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

This article describes how LendingTree built a production multi-agent mortgage assistant on Amazon Bedrock to help consumers navigate complex home-buying decisions with accurate information and personalized guidance.

• Three-agent architecture: Supervisor orchestrator coordinating Education and Matching worker agents using LangGraph and Model Context Protocol
• Amazon Bedrock Guardrails provide content filtering, PII redaction, and prompt threat screening for regulatory compliance
• Education worker maintains specialized Knowledge Bases with semantic chunking for mortgage concepts; Matching worker connects users with personalized lending options
• Multi-model approach uses Amazon Nova Pro for complex reasoning and Nova Lite for conversational responses to balance reliability and cost
• Conversation memory persists across agent handoffs using PostgreSQL checkpointer on Amazon RDS
• Production results: 97% containment without human escalation, average 6.2 messages per conversation, engaged users sustain 10+ message sessions
• Key learnings include semantic chunking for retrieval quality, parallel safety checks for latency, and task-based model routing for cost optimization

The system demonstrates that AI agents can serve consumers in regulated industries at scale while maintaining strict compliance standards through thoughtful architecture and safety-first design.