Intelligent Expense Chatbot: Agentic AI Teaching Practice Driven by a Multi-Agent Dialogue System

This article introduces an intelligent expense chatbot system based on a multi-agent architecture, which teaches generative AI and Agentic AI concepts through real financial management scenarios and demonstrates the innovative application of multi-objective dialogue systems in the field of financial intelligence. This system is not only a technical practice but also a teaching case that combines theory and application, helping learners understand the autonomous decision-making, tool calling, and multi-agent collaboration mechanisms of Agentic AI.

In recent years, the AI field has developed rapidly from traditional machine learning to deep learning and then to generative AI. However, a single generative model lacks autonomous decision-making and tool-calling capabilities when handling complex tasks. Agentic AI emphasizes that the system can autonomously plan, use tools, and interact to complete multi-step tasks like an intelligent agent. Based on this concept, this project builds a teaching practice system, allowing learners to intuitively understand the multi-agent collaboration mechanism through life-like financial management scenarios.

Core Idea of Multi-Agent Architecture

Unlike traditional single large model dialogue systems, this system adopts a multi-agent architecture where different AI agents are responsible for specific subtasks and collaborate to complete complex requirements. The division of labor is similar to that of a human team, and information is exchanged through standardized communication mechanisms.

Division of Agent Roles and Responsibilities

• Intent Understanding Agent: Parses users' natural language input, identifies real intentions (such as querying expenses, setting budgets, analyzing consumption habits), and triggers corresponding processes.
• Data Retrieval Agent: Interacts with the database to query financial information such as transaction records and account balances, and converts vague descriptions into precise retrieval conditions.
• Analysis and Reasoning Agent: Undertakes data analysis tasks, identifies consumption patterns, calculates expenditure proportions, predicts trends, and may call statistical tools or machine learning models.
• Response Generation Agent: Converts the results from other agents into natural and friendly dialogue responses, and ensures coherence by integrating context.

Materialization of Generative AI Concepts

Beginners' understanding of generative AI often stays on the surface. This project demonstrates how generative AI understands intentions, generates context-aware responses, and maintains the coherence of multi-turn dialogues through actual dialogue interactions, allowing learners to see the complex language processing behind simple expense queries.

Demonstration of Core Features of Agentic AI

• Autonomous Planning Capability: When handling complex requests, it first decomposes tasks (e.g., analyzing last month's consumption and giving suggestions will be broken down into steps like data query, analysis, and suggestion generation).
• Tool Calling Mechanism: Shows when the AI calls external tools (database, calculator, etc.) and how it processes the returned results, embodying the "think-act-observe" cycle.
• Multi-Objective Coordination Capability: Coordinates the work priorities of different agents, properly handles multiple user needs while maintaining smooth dialogue.

Personal Financial Management Assistant

Ordinary users do not need to learn complex accounting software; they can get financial analysis and suggestions by describing their needs in natural language.

Enterprise Training Tool

Provides enterprises with a low-threshold experimental platform, allowing employees to understand the principles of Agentic AI through interaction and lay a foundation for subsequent complex AI application development.

Education and Training Case

As a case combining theory and practice in AI courses of universities and training institutions, students can deeply understand technical details by reading source code, modifying parameters, and expanding functions.

Agent Communication Protocol

Design a clear communication protocol, define message formats, state transfer, and error handling mechanisms to ensure efficient collaboration between agents.

Context Management

Maintain dialogue state, track entities mentioned by users (specific accounts, time periods), and correctly reference this information in subsequent interactions.

Fault Tolerance and Recovery

When an agent fails to process a task, ensure system robustness through retries, downgrading to alternative solutions, or explaining the situation to the user and requesting clarification.

With the improvement of large language model capabilities, the application scenarios of Agentic AI will become more extensive. This project can be further expanded: integrate more types of agents (prediction, recommendation agents), support more complex financial planning tasks, and even link with smart home and health monitoring systems to provide comprehensive life services. Through this project, learners can master the core concepts of Agentic AI and cultivate practical abilities to design and implement complex AI systems.