LoanPro Deploys Multi-Agent Loan Servicing Assistant, Cutting Call Time by Up to 15%
Summary:LoanPro, working with the AWS Generative AI Innovation Center and AllCloud, has deployed a multi-agent assistant for mortgage servicing teams on Amazon Bedrock. Built with the Strands Agents SDK, the system separates account retrieval, account summarization, next-best-action recommendation, and response composition into specialist agents. Amazon Bedrock AgentCore Runtime and Gateway connect those agents to LoanPro’s production APIs through the Model Context Protocol, while API Gateway, Lambda, a
A Controlled Workflow, Not a Chat Window
LoanPro has put an AI assistant into production for mortgage servicing teams. The system does not replace service agents outright. Instead, it compresses payment records, loan terms, support tickets, and workflow queues into an account briefing and recommends a next action.
Its central contribution is not a single foundation model. It is an execution chain combining AWS managed services, the Strands Agents SDK, MCP tool access, and LoanPro’s existing authorization and compliance systems.
Request Path: From Public API to Production Data
The implemented path is straightforward but operationally significant:
- Requests enter LoanPro API v2 through Amazon API Gateway.
- A handler Lambda places each request on Amazon SQS, decoupling traffic spikes from agent execution.
- An Agent Interceptor Lambda invokes Amazon Bedrock AgentCore Runtime.
- A Composer Agent routes the request to specialists for account retrieval, summarization, or next-action recommendation.
- Specialist agents discover and invoke MCP tools through AgentCore Gateway.
- The Loan Summary Tool Lambda validates the caller’s signature, then retrieves data from LoanPro’s loan-management system running on EC2 behind a Network Load Balancer.
- LoanPro applies its own RBAC and compliance checks to every request and writes actions to an audit trail.
The tool catalog is registered in AWS Systems Manager Parameter Store. A separate Invoke API Tool can display information panels or start predefined automation sequences; it is not an unrestricted interface for executing arbitrary internal operations.
What Multi-Agent Separation Buys—and What It Costs
Rather than asking one model to retrieve data, summarize it, choose an action, and generate a response, the system assigns those tasks to separate agents.
The main benefit is evaluability. Retrieval failures, factual summarization errors, and inappropriate action recommendations can be measured independently rather than collapsed into an opaque end-to-end score. The modular design also makes it easier to add agents for collections and loan origination without redesigning the entire platform.
The cost is additional context transfer, model invocation, and orchestration latency. If the Composer passes a lossy natural-language summary to another agent, important fields may disappear and downstream errors can compound. A production implementation should therefore retain structured account state, field-level evidence, and tool-call traces instead of repeatedly paraphrasing summaries between agents.
The article says the system can use Claude Sonnet 5.5, Haiku 5.5, and Opus 5.5 through Bedrock, but it does not disclose the exact routing policy. A sensible design would use a lighter model for classification and field extraction, reserve stronger models for difficult synthesis or recommendations, and escalate based on confidence, risk, and latency budgets.
MCP Is an Interface, Not a Security Boundary
AgentCore Gateway exposes LoanPro APIs through MCP, centralizing tool discovery, connectivity, and authentication. Adding a new tool does not require redeploying every agent.
MCP itself, however, only standardizes how tools are described and invoked; it does not enforce authorization. LoanPro’s server-side checks of SSO signatures, user permissions, and compliance rules are the more important control layer. Tool descriptions can also become prompt-injection surfaces, so authorization must be enforced server-side by user, account, field, and action rather than relying on the model to select the correct arguments.
The Data Retrieval Tool spans eight API endpoints and touches roughly 1,400 fields per account. Broad context improves completeness, but it also increases token cost, latency, privacy exposure, and the chance that an irrelevant field contaminates the summary. Production systems should use field allowlists, task-specific retrieval, and minimum-necessary context instead of placing all 1,400 fields into the prompt.
Managed Infrastructure Reduces Work—and Adds Lock-In
Using Bedrock, AgentCore, and Lambda removes the need to operate model endpoints, the agent runtime, and autoscaling infrastructure. Lambda and SQS can also absorb request bursts. LoanPro’s separate proof of concept for converting natural language into Clojure rule code suffered from timeouts, missing monitoring, and no deployment strategy, illustrating how much of “productionization” lies outside the model itself.
The trade-off is vendor dependency, regional availability, and model-version coupling. Claude model availability differs across AWS regions, and an upgrade can alter summary structure, recommendation style, or tool selection. Teams should pin model and prompt configurations, run regression tests for every version change, and retain a deterministic fallback path that leaves agents with account summaries and manual controls when a model is unavailable.
SQS provides decoupling and burst absorption, not exactly-once execution. Tool operations should use idempotency keys, bounded retries, a dead-letter queue, timeouts, circuit breakers, and correlation IDs. Without those controls, redelivery can produce duplicate payment plans, notifications, or inconsistent account states.
How to Interpret the 99% Accuracy and 15% Improvement
The team evaluated retrieval, summarization, and action selection on 150 accounts labeled by subject-matter experts and reports that the system met a 99% accuracy target. This is more useful than a demo, but the public description does not define the metric precisely: account-level exact match, field-level accuracy, or recommendation acceptability? Were false negatives on high-risk actions counted? Was there overlap between development and evaluation data?
The production figures also need careful interpretation:
- Account summaries are reportedly produced in under 15 seconds rather than through three to five minutes of manual navigation, but no P50, P95, or P99 latency is disclosed.
- The customer-call-time reduction of up to 15% comes from an AllCloud case study and uses an upper-bound formulation.
- The nearly 8% decline in internal tickets may also reflect process changes or differences in the sampled lender population.
- For more than 800 servicing professionals at early launch partners, saving as much as nine minutes per agent-hour would imply roughly 120 agent-hours of capacity per operating hour. That is an upper-bound arithmetic projection, not an independently measured productivity gain.
A stronger evaluation would report retrieval recall, factuality for critical fields, unsupported-recommendation rate, false-action rate on high-risk operations, end-to-end P95 latency, token cost per account, and human-override rates, stratified by loan state, issue type, and agent experience.
A Deployment Sequence for Regulated Workflows
- Start read-only, then advise, then execute. Launch account summaries first, followed by low-risk recommendations. Keep human confirmation for payment plans, deferrals, and escalations, with parameter-level policies for every action.
- Enforce permissions at the tool boundary. Use SSO/JWK, least-privilege scopes, account-level authorization, and short-lived credentials. Logs must identify the human initiator, model, tool, arguments, result, and accountable approver.
- Build a replayable evaluation set. Extend the 150 labeled cases with delinquency, dispute, refund, identity, and high-risk action scenarios. Preserve the real case distribution and use temporal splits to prevent account-level leakage.
- Control context size. Apply task-level retrieval, field redaction, and summary caching across the 1,400-field account surface. Retain references to source evidence so a plausible but unsupported summary cannot hide missing data.
- Complete the AgentOps layer. Trace composer routing, every model call, tool latency, queue wait, token cost, Guardrails triggers, and human rewrites. Aggregate response quality alone will not locate the bottleneck.
- Adopt tiered autonomy. Permit automatic execution only for low-risk, reversible operations. Changes to interest, repayment obligations, or credit status should require a named human approver. Bedrock Guardrails can filter content and validate grounding, but they do not replace a deterministic business-rules engine.
LoanPro plans to extend the architecture into collections and origination. The reusable asset is not a prompt but the combination of account-state modeling, specialist boundaries, an MCP tool catalog, server-side authorization, an audit chain, and continuous evaluation. Keeping that control layer deterministic allows the number of models and agents to grow; replacing it with model judgment would make business risk scale with automation.
— END —Source: AWS Machine Learning Blog (2026-11-17)
Tags: #Amazon Bedrock AgentCore #Strands Agents #Multi-Agent Systems #Model Context Protocol #Loan Servicing
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments