이미지: METAL LAB 생성
Summary
- LendingTree has built a mortgage assistant on Amazon Bedrock made up of three agents: a supervisor and dedicated education and matching agents
- The system combines LangGraph, the Model Context Protocol (MCP), and Amazon Nova models, with Amazon Bedrock Guardrails handling content filtering and PII masking
- Bedrock AgentCore launched generally after the service was already in production, so the system was deployed as containers on Amazon ECS and AWS Fargate instead
- 구성
- 슈퍼바이저 1개 + 워커 에이전트 2개(교육, 매칭)
- 오케스트레이션
- LangGraph 기반 plan-and-execute 상태 머신
- 연동 프로토콜
- Model Context Protocol(MCP)
- 배포 환경
- Amazon ECS + AWS Fargate 컨테이너
- 안전 장치
- Amazon Bedrock Guardrails로 유해 콘텐츠 필터링·PII 마스킹·프롬프트 위협 탐지
LendingTree has built a multi-agent assistant on Amazon Bedrock to help consumers navigate mortgage decisions, according to the AWS ML blog. The goal is to let consumers work through complex choices — purchase versus refinance, government-backed loans, loan terms, and rate structures — through natural conversation.
Three agents split the workload
The system consists of a supervisor agent and two worker agents, one for education and one for matching. The supervisor is built as a LangGraph-based state machine that orchestrates the overall flow using a plan-and-execute approach, while each worker handles its assigned domain. The Model Context Protocol (MCP) is used to connect the agents, and the foundation models are reportedly Amazon Nova models served through Amazon Bedrock.
Dual safeguards for compliance
Given the regulatory nature of the financial industry, content filtering, protection of personally identifiable information (PII), and compliance oversight were essential requirements. User inputs and model outputs pass through Amazon Bedrock Guardrails, which handles detection of hate speech and profanity, PII masking, and prompt-threat screening. In addition, an LLM-based safety classifier checks conversations for policy compliance in parallel, providing dual verification without adding latency.
Srinivas Madabushi, head of technology at LendingTree, said, "Bedrock's built-in guardrails let us launch a service with security and compliance built in from the start."
The entire service runs as containers on Amazon ECS and AWS Fargate, including a React-based chat interface. LendingTree reportedly kept its existing ECS setup rather than moving to Bedrock AgentCore, since AgentCore only became generally available after the system was already deployed in production.
