2 Working with ADPs: Building a mortgage advisor tool

This chapter uses a mortgage advisor as a practical example for assembling agent design patterns into a reliable, regulated, and tool-using system. Mortgage advice is a strong case study because the agent must ask targeted questions, rely on specialized and time-sensitive knowledge, perform accurate long-term calculations, and avoid producing language that could violate financial regulations. Rather than treating the problem as a single conversational workflow, the design breaks the task into distinct stages that can be composed cleanly.

The overall flow begins with an information-gathering phase, where an interlocutor actively collects the missing details needed for the user’s request and stores them in structured form. Once enough information is available, a supervised team of specialist analysts takes over: each analyst is implemented as a ReAct-style agent with access to relevant tools, and a supervisor coordinates them to produce an aggregate analysis. The chapter emphasizes that this division into subagents helps manage complexity, keeps contexts smaller, and makes each specialist easier to improve and test independently.

After the analysis is complete, a reporting component turns the results into a user-facing mortgage advisory draft, and a compliance guardrail checks whether the language is acceptable under financial regulations. If the report fails compliance, feedback is fed back into the writing step so the system can refine the output, with retries capped to prevent endless looping. The result is a modular agent architecture in which runtime, orchestration, analysis, writing, and safety checks are separated but coordinated through shared state, showing how ADPs can support dependable agent behavior end to end.

• Agent operations can be visualized using the agent flow diagrams (AFDs) introduced in this chapter.
• To implement an agent like our example Mortgage Advisor agent, we compose high-level patterns to cleanly isolate the individual pieces of its functional logic.
• It can be advantageous to use a variety of patterns in a single application, from the human-in-the-loop interlocutor, through a simple centralized team of (sub)agents, to a refiner and guardrails. These patterns all contribute to the user experience and agent-response quality in different ways.
• One component of an agent’s implementation may have the characteristics of multiple patterns when appropriate. Although our MA example implementations are simple, the components implementing each pattern can be considerably more complex while still satisfying the same defining characteristics.

FAQ