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

Shopping for a house is without doubt one of the greatest monetary selections most individuals face, and LendingTree constructed a multi-agent mortgage assistant on Amazon Bedrock to make the method extra simple. The assistant educates debtors, understands their scenario, and offers tailor-made choices in a pure dialog. Debtors should weigh buy or refinance, standard or government-backed, 15-year or 30-year phrases, and glued or adjustable charges. On prime of that, there’s jargon like “low cost factors,” “origination charges,” and “debt-to-income ratio.” It’s no marvel many individuals really feel misplaced earlier than they even begin.

LendingTree has been serving to shoppers type by way of these selections for over 25 years, connecting tens of millions of individuals with lenders to seek out aggressive mortgage affords. The corporate was constructed on a perception that everybody deserves the instruments and data to make assured monetary selections. The AI-powered mortgage assistant is that subsequent step.

The answer needed to meet the identical requirements which have guided LendingTree from day one: correct info, clear steering, and rigorous safety of person information. Working throughout the regulatory necessities of the mortgage trade means content material filtering, personally identifiable info (PII) safety, and compliance oversight aren’t non-obligatory options. They’re non-negotiable. That made Amazon Bedrock and its built-in guardrails the correct basis.

“Our purpose was to be a trusted accomplice within the home-buying journey, a information that educates shoppers, understands their scenario, and matches them with the correct provide. The inspiration fashions and built-in guardrails in Amazon Bedrock allow us to ship that with safety and compliance from the beginning.”

— Srinivas Madabushi, SVP Know-how, LendingTree

Many corporations within the trade have added chatbots for fundamental questions. LendingTree needed to go additional, answering the laborious questions and matching debtors with aggressive affords. That took multiple agent.

Resolution overview

To deal with these challenges, LendingTree deployed three unbiased AI brokers: a supervisor and two specialised staff (schooling and matching), coordinated utilizing LangGraph, the Mannequin Context Protocol (MCP), and basis fashions powered by Amazon Bedrock. All brokers run containerized providers on Amazon Elastic Container Service (Amazon ECS) with AWS Fargate.

LendingTree selected Amazon Bedrock for its multi-model flexibility and inherited AWS governance controls, which their compliance crew required. The answer was deployed on Amazon ECS as a substitute of Amazon Bedrock AgentCore as a result of it was already in manufacturing when AgentCore reached basic availability. AgentCore is a platform to construct, join, and optimize brokers at scale, with any framework or mannequin.

The next determine reveals the high-level structure.

LendingTree multi-agent structure for mortgage steering

(1) The patron-facing chat interface is a React utility operating as a containerized service on Amazon ECS with Fargate, serving shoppers on net and cellular browsers by way of a public endpoint.

(2) Person inputs and mannequin outputs go by way of Amazon Bedrock Guardrails for content material filtering (reminiscent of hate and profanity detection) and PII redaction. Incoming messages are additionally screened for immediate threats.

In parallel with Guardrails, a security classifier based mostly on a big language mannequin (LLM) enforces LendingTree’s conversational coverage. The 2 security checks run concurrently, so this added assurance doesn’t add latency.

A business-logic layer handles LendingTree’s operational guidelines: routing complicated points to human assist and redirecting off-topic conversations.

(3) The Supervisor agent is the orchestrator, constructed on LangGraph as a state machine that follows a plan-and-execute sample. It really works like a well-organized crew: the Supervisor understands the large image and decides what must occur, whereas the employees deal with what they’re every greatest at.

The Supervisor is a graph of nodes and edges. Nodes do the work (intent evaluation, execution planning, response composition), and edges determine what occurs subsequent based mostly on the end result (path to Schooling, path to Matching, or reply straight). This makes the movement express: each path by way of the system is outlined, auditable, and traceable.

The Supervisor analyzes person intent with Amazon Nova Professional and produces an execution plan. It then routes to the suitable employee over MCP utilizing connection pooling for low-latency calls. For mannequin availability by Area, see Supported fashions by AWS Area in Amazon Bedrock.

Not each activity wants probably the most highly effective mannequin. The Supervisor implements multi-model structure, routinely choosing between Amazon Nova Professional (for complicated reasoning and significant classification) and Amazon Nova Lite (for conversational responses and light-weight classification) relying on what the duty calls for. This balances reliability with price.

Planning and execution are separated by design. When a dialog goes flawed, the crew can hint precisely which node made which choice.

(4) The Schooling employee is the affected person educator. It helps customers perceive mortgage ideas, explains product variations, and offers steering by way of interactive dialog. It runs its personal LangGraph workflow and maintains specialised Amazon Bedrock Data Bases, the absolutely managed retrieval augmented technology (RAG) functionality, tailor-made to its area, backed by Amazon OpenSearch Service because the vector retailer. Each response is grounded in actual paperwork slightly than counting on mannequin data alone.

(5) The Matching employee is the connector. It gathers person preferences and calls LendingTree’s inner provide, eligibility, and fee APIs. From there, it delivers customized lending choices based mostly on the person’s precise scenario. It interprets qualification standards and helps customers examine choices facet by facet.

(6) Shared providers deal with dialog reminiscence and session state by way of the LangGraph PostgreSQL checkpointer on Amazon Relational Database Service (Amazon RDS). This lets conversations persist throughout turns, agent handoffs, and repair restarts. Customers can pause, come again later, or ask follow-ups with out shedding context.

(7) Amazon Bedrock offers a broad set of basis fashions suited to a variety of use circumstances. These vary from complicated reasoning and significant classification (Amazon Nova Professional) to cost-efficient conversational responses and light-weight classification (Amazon Nova Lite), all accessible by way of a single API. Constructed-in Guardrails ship configurable security controls, whereas Amazon Bedrock Data Bases join brokers to authoritative content material by way of vector shops reminiscent of Amazon OpenSearch Service for hybrid retrieval.

(8) Inner LendingTree providers and APIs present lender search, fee and product information, eligibility and prequalification checks, provide comparability, and person profile info.

(9) All providers run on Amazon ECS with AWS Fargate, the place every agent scales independently by itself demand indicators. Infrastructure is managed by way of Terraform and deployed by way of GitLab steady integration and steady supply (CI/CD) pipelines with automated testing and well being checks.

Following a dialog finish to finish

Suppose a person asks, “What’s the distinction between FHA and traditional, and which one suits me?”

The message flows by way of Amazon Bedrock Guardrails (2) for security screening. The Supervisor agent (3) then hundreds dialog historical past from dialog reminiscence (6) and analyzes intent, recognizing two wants: an evidence and a personalised comparability. It routes the academic query to the Schooling employee (4) over MCP, and the matching query to the Matching employee (5). Every employee operates independently and returns its end result. The Supervisor aggregates each into one coherent reply, operating the output again by way of Guardrails earlier than delivering it. The total trade is checkpointed to PostgreSQL (6), so the subsequent flip picks up with full context.

To the person, it’s one steady dialog. Behind the scenes, it’s a coordinated workflow the place each choice level is traceable.

Deployment and operations

As a result of the brokers talk by way of MCP, every could be up to date, scaled, and rolled again independently. The Schooling employee can ship a brand new data base with out touching the Supervisor or the Matching employee. Operationally, the problem distinctive to multi-agent techniques is tracing a single dialog that spans a number of providers. Amazon CloudWatch logs and AWS X-Ray distributed tracing permit the crew to observe a single dialog’s journey throughout all three brokers, with per-agent metrics and detailed timing.

Outcomes

The assistant has been dealing with actual mortgage conversations in manufacturing since late 2025. The metrics on this part are based mostly on LendingTree’s inner manufacturing analytics from launch by way of Q1 2026, as reported by the LendingTree engineering crew. Throughout that interval, it served roughly 1,960 conversations and 12,100 messages, averaging 6.2 messages per trade. These aren’t one-shot lookups: engaged customers maintain multi-turn periods averaging 10+ messages over 9 minutes, asking follow-up questions, refining their scenario, and exploring choices conversationally.

From schooling to motion. The dialog information reveals how shoppers truly use an AI mortgage assistant. Early within the rollout, 75% of conversations had been academic. Customers requested questions like “What’s an FHA mortgage?”, “How does a HELOC work?”, or “What credit score rating do I would like for a standard mortgage?” Because the system matured and phrase unfold, transactional intent grew: over 50% of current conversations now contain fee comparisons, lender matching, or prequalification, actions tied on to conversion.

Dialog depth indicators belief. The typical session size of 6.2 messages tells one story. The engaged-user common of 10+ messages over 9 minutes tells a special one. Customers don’t maintain that sort of engagement with a system they don’t discover helpful. The Schooling employee builds understanding, and customers naturally progress to the Matching employee after they’re able to act. No compelled handoff, no “begin over” second.

The questions reveal actual complexity. The most typical subjects embrace mortgage sort comparisons (FHA, standard, and VA), qualification standards for particular credit score profiles, fee negotiation methods, closing timelines, and down fee trade-offs. These are precisely the situation-dependent questions that static FAQ pages can’t reply effectively. They require context: what the person mentioned three turns in the past, what they qualify for, what they’re optimizing for. That is the place dialog reminiscence and the Supervisor’s intent evaluation earn their complexity, and the system maintains the thread so every reply builds on the final.

Containment. Over 97% of conversations had been dealt with end-to-end with out human escalation, and solely about 3% of customers explicitly requested a dwell agent. For a regulated monetary product the place the questions are genuinely laborious and getting it flawed has actual penalties, that containment fee is important. It demonstrates the system’s capability to function as a self-contained advisory service slightly than a triage layer in entrance of a name middle.

What the information confirms. The engagement depth is the sign that issues most. Higher-informed customers make higher selections. A conversational interface surfaces preferences and constraints {that a} static kind by no means captures together. Particulars like “I’m a veteran with a 650 credit score rating trying to purchase in Colorado Springs within the subsequent 30 days” make downstream lender matches extra related. That’s finally what serves the patron.

What LendingTree realized, and what’s subsequent

Getting a multi-agent system to manufacturing taught the crew as a lot about structure as about any single function.

After transport the mortgage assistant, LendingTree realized that constructing one agent is easy, however scaling to many brokers requires shared foundations. Relatively than constructing each new function as a one-off, LendingTree is investing in reusable items. These embrace shared context layers for information entry, MCP contracts for instrument integration, and standardized deployment processes. A registry lets groups uncover present capabilities as a substitute of reinventing them. This lets groups construct smaller, domain-focused capabilities that may be composed into broader shopper experiences. Constant requirements for security, observability, analysis, and reuse tie all of it collectively.

Agent design learnings:

• Semantic chunking for Data Bases. Breaking paperwork into semantically coherent chunks slightly than fixed-size chunks improved retrieval high quality considerably, as a result of chunk boundaries align with pure subject breaks.
• Data Base battle decision. With a number of KBs, contradictory info generally surfaced. Area-based filtering and supply prioritization solved this: inner LendingTree content material takes priority for product-specific questions, whereas exterior assets serve basic mortgage schooling.
• Inter-agent context passing. Employee brokers initially lacked consciousness of the broader dialog. Passing full dialog historical past and intent summaries in every MCP request gave staff the context they should present related responses.
• Question rewriting. Quick person responses like “undecided” or “sure” get rewritten into significant, searchable queries utilizing dialog historical past earlier than retrieval. This considerably improved retrieval high quality.
• Guardrail tuning is ongoing work. Early configurations blocked legit questions as a result of mortgage terminology tripped content material filters. Tuning towards real looking dialog information resolved this.
• Activity-based mannequin routing saved prices in examine: Nova Professional solely the place reasoning demanded it, Nova Lite in all places else.

These are agent design issues, not infrastructure issues. They continue to be essential no matter runtime alternative, and characterize the lasting data from this implementation.

Infrastructure learnings:

• Dialog state administration. Early variations misplaced context throughout agent handoffs. A unified PostgreSQL-backed checkpointer with express state serialization solved this, permitting conversations to persist throughout agent transitions and system restarts.
• Operating security checks in parallel preserved latency with out weakening safety.
• Unbiased agent scaling required hand-wiring ECS containers, well being checks, and deployment pipelines per agent.

The crew is now evaluating re-architecting onto AgentCore to dump undifferentiated infrastructure work. That might keep away from hand-wiring PostgreSQL checkpointers, ECS containers, and agent scaling, letting engineering effort shift from “maintain the runtime alive” to “maintain optimizing the agent logic and area data.”

Conclusion

LendingTree’s manufacturing multi-agent system demonstrates that AI brokers can serve shoppers in closely regulated industries, not solely as prototypes, however as manufacturing techniques operating 24/7. Amazon Bedrock (Nova Professional and Nova Lite, Data Bases, and Guardrails) supplied the mannequin and security basis. LangGraph dealt with agent orchestration, and MCP dealt with agent-to-agent communication. Collectively, they ship customized mortgage steering at scale whereas sustaining strict compliance requirements.

For organizations constructing multi-agent techniques: separate planning from execution so the system is debuggable. Make security a structural pillar slightly than a skippable step. Design from day one for the reusable capabilities that may allow you to scale the sample throughout your group.

To get began with multi-agent techniques on AWS, discover the Amazon Bedrock documentation and Amazon Bedrock AgentCore for managed runtime capabilities.

Concerning the authors