Agentic AI for Title Operations Workflow Optimization

rocket · Finance · 2026 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Mortgage title operations

Rocket Close needed to reduce title-operation bottlenecks from state-specific exams, fragmented systems, and manual research.

Constraints:

Identity/token validation
Title-policy knowledge
MCP tool execution
Operational accuracy

Architecture

Rocket Close Supercharger is a Strands Agents and Bedrock title-operations assistant. A user opens a WebSocket connection with a JWT, identity is validated, a domain-specific Strands agent receives the prompt, queries Bedrock Knowledge Bases for policies and procedures, selects MCP functions, retrieves order information from Atlas Web API, synthesizes order-specific context, and streams a role-specific response. Bedrock Guardrails, row-level entitlements, audit logging, monitoring, and clear separation between Strands orchestration and MCP tools protect regulated title operations.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  operations-user[Operations user] --> websocket-interface[WebSocket interface]
  websocket-interface --> token-validation[Token validation]
  token-validation --> strands-title-agent[Strands title agent]
  strands-title-agent --> bedrock-knowledge-bases[Bedrock Knowledge Bases]
  strands-title-agent --> mcp-tools[MCP tools]
  mcp-tools --> atlas-web-api[Atlas Web API]
  bedrock-knowledge-bases --> context-synthesis[Context synthesis]
  atlas-web-api --> context-synthesis
  context-synthesis --> websocket-interface
  guardrails-and-entitlements[Guardrails and entitlements] --> strands-title-agent

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Operations user`	Rocket Close team member asking title-operation questions.	`WebSocket interface`
`WebSocket interface`	Streaming request/response channel authenticated with JWT.	`Operations user`, `Token validation`
`Token validation`	Identity provider and Istio validation for user access.	`WebSocket interface`, `Strands title agent`
`Strands title agent`	Domain-specific agent driving title-operation conversation and tool selection.	`Token validation`, `Bedrock Knowledge Bases`, `MCP tools`
`Bedrock Knowledge Bases`	Policy and procedure retrieval for title exams and operations.	`Strands title agent`, `Context synthesis`
`MCP tools`	Distinct tool interfaces for operational data sources.	`Strands title agent`, `Atlas Web API`
`Atlas Web API`	Order-information API used by MCP tools.	`MCP tools`, `Context synthesis`
`Context synthesis`	Combines policy, procedure, and order-specific context into a response.	`Bedrock Knowledge Bases`, `Atlas Web API`, `WebSocket interface`
`Guardrails and entitlements`	Bedrock Guardrails, row-level entitlements, logging, monitoring, and audit trails.	`Strands title agent`

Data flow

A user connects through WebSocket with a JWT token. The identity provider validates access, the exam title agent is invoked, the agent retrieves relevant policy and order context from a knowledge base and MCP tools, the Atlas Web API supplies order information, the system synthesizes context, and the response streams back through WebSocket.

3 supporting quotes

“WebSocket handshake – The user starts a connection through an HTTP request with a JWT token. 2. Token validation – The identity provider validates the token through Istio and establishes a WebSocket connection. 3. Exam title agent invocation – The Strands”article, stated

“MCP tool execution – MCP tools process the request, retrieving order information from the Atlas Web API. 7.”article, stated

“Response delivery – The combined response streams back to the user through WebSocket. 9. Response Rendering – The synthesized response is progressively streamed back to the Chatbot UI.”article, stated

Why this architecture

Rocket chose Strands plus MCP because title operations need flexible natural-language orchestration over many data sources without hardwiring every workflow step. MCP tools make data sources extensible and single-call retrieval faster, WebSocket streaming improves perceived performance, and entitlements/guardrails keep customer-sensitive order data bounded.

3 supporting quotes

“each data source is exposed as a distinct tool that Strands Agents can invoke. This approach delivers three advantages: * Extensibility – New data sources can be added as additional tools without restructuring the core architecture. The team made this design”article, stated

“MCP tools retrieve the necessary order information in a single call before using LLM synthesis to extract relevant details, alleviating the need for multiple database queries. This architectural philosophy extended to maintaining a clear separation of concerns”article, stated

“The team found that WebSocket-based streaming delivered immediate user feedback, improving perceived performance even when handling complex queries.”article, stated

Models used: Anthropic Claude through Amazon Bedrock

Scale

Supercharger reduced contact-center calls and emails by 30%.
“The solution delivered immediate operational efficiency gains for the operations and client relations teams, reducing the number of incoming calls and emails to the contact center by 30% through its question-answering capability.”article, stated
Prompt and architecture refinement achieved 3x latency improvements.
“performance was optimized through architectural refinement and better prompting techniques that reduced the number of calls the agent made to the LLM, achieving 3x latency improvements and reduced costs.”article, stated

Alternatives considered (roads not taken)

Multiple database queries before synthesis

Rejected because: A single MCP order-information call improved retrieval performance.

“call before using LLM synthesis to extract relevant details, alleviating the need for multiple database queries. This architectural philosophy extended to maintaining a clear separation of concerns between Strands Agents and MCP tools, creating a flexible foundation capable of evolving alongside”article, stated

Prompt the agent with deterministic step-by-step workflows

Rejected because: Goal-oriented prompting let the agent orchestrate dynamically and remain maintainable.

“The team learned that effective LLM prompting focuses on describing what the agent should accomplish rather than prescribing how, because removing deterministic steps allowed the agent to orchestrate dynamically using its inherent capabilities, proving more adaptable than custom approaches.”article, stated

Patterns

`tool-use-orchestration` (title-ops)

The agent selects functions and invokes MCP tools to retrieve operational data.

“4. Knowledge base query – The agent searches the knowledge base for relevant policies and procedures. 5. Tool selection – The agent determines which function to invoke and with which parameters.”article, stated

`rag-retrieval` (knowledge-base)

The title agent searches a knowledge base for policy and procedure context.

“4. Knowledge base query – The agent searches the knowledge base for relevant policies and procedures.”article, stated

Tradeoffs

Let the agent orchestrate dynamically instead of scripting every step.

Chose Goal-oriented prompting and agent-driven orchestration over Prescribing deterministic step-by-step workflows inside prompts.

Rationale: Rocket found that removing deterministic steps made the system more adaptable and maintainable.

How to keep title-agent retrieval fast.

Chose single-call MCP order retrieval before LLM synthesis over multiple database hops inside the agent loop.

Rationale: Reducing retrieval round trips improved latency while keeping Strands focused on orchestration and synthesis.

“efficient data retrieval stands as a cornerstone of performance”article, stated

Gotchas (what bit them)

Title-operation agents needed fast retrieval and clean tool contracts to stay maintainable.

Root cause: Multiple database hops, over-prescribed prompts, and weak tool descriptions would slow and constrain the agent.

Resolution: Rocket used MCP tools for single-call retrieval, separated agent/tool concerns, and described goals rather than deterministic steps.

“The experience revealed that efficient data retrieval stands as a cornerstone of performance, leading them to architect a streamlined solution where MCP tools retrieve the necessary order information in a single call”article, stated

Evaluation

The post reports operational metrics including contact-center inquiry reduction and latency improvement.

Metrics: contact center inquiries · latency

“ural language interactions. The solution delivered significant operational improvements including a 30% reduction in contact center inquiries, enhanced state exam accuracy through real-time insights, improved client satisfaction through automation of routine tasks, and 3x latency improvements throug”summary, stated

Outcomes

The summary reports a 30% inquiry reduction and 3x latency improvement.

Vendor framing is high, so the outcomes should be treated as reported claims.

“y through real-time insights, improved client satisfaction through automation of routine tasks, and 3x latency improvements through architectural optimization.”summary, stated