Ticket Subject Agent: An Introductory Practice of Agentic Workflow Based on Next.js and Vercel AI SDK

This article introduces an example project of Agentic Workflow built with Next.js and Vercel AI SDK—Ticket Subject Agent, which demonstrates how large language models can implement intelligent querying and analysis of ticket data through tool calls. It is a concise introductory case for understanding the AI Agent architecture. Focusing on the ticket data analysis scenario, the project transforms AI from a passive responder to an active task executor by clearly defining tools and workflows, helping developers grasp the core concepts of Agentic Workflow.

In traditional AI applications, the interaction between users and models is a direct question-and-answer mode, which has limitations such as inability to actively obtain external data, perform specific operations, or complete tasks according to processes. Agentic Workflow (intelligent agent workflow) solves these problems: large language models become intelligent agents that make autonomous decisions, call tools, and execute tasks—they can understand intentions, select tools, and complete complex tasks step by step.

Ticket Subject Agent is a deliberately concise introductory example, not a complete customer service system. It demonstrates the principles of intelligent agent workflow through the ticket data analysis scenario. The tech stack includes: Next.js (a React full-stack framework, convenient for development and deployment), Vercel AI SDK (an AI application development toolkit that simplifies model calls and streaming responses), and CSV dataset (ticket data from February 2026 as the data source).

From Black Box to Controllable: General chatbots cannot control step order, reusable tools, result display, safety guardrails, or structured output. Agentic Workflow achieves controllability by clearly defining tools and workflows.

Tool-Driven: The model is equipped with four tools:

1. getTicketOverview: Obtain an overview such as total number of tickets and status distribution;
2. searchTickets: Multi-dimensional ticket retrieval;
3. findContactsWithMultipleTickets: Find contacts who submitted multiple tickets;
4. findLikelyDuplicateSubjects: Discover tickets with suspected duplicate subjects.

Role of Vercel AI SDK: Streaming response processing (real-time display of replies), tool call abstraction (automatically handles tool selection/parameter parsing/result passing), multi-model compatibility (supports OpenAI, etc.).

Tool Definition Pattern: Tools are defined using TypeScript + Zod, including descriptions, parameter validation, and execution logic, ensuring type safety, self-description, and testability.

Workflow Orchestration: User input → Intent understanding → Tool selection → Parameter extraction → Execution call → Result integration → Output display. It can be executed cyclically (continue if one tool call is insufficient).

Educational Value: Realistic scenario (ticket management is common), moderate code volume (focuses on core concepts), modern tech stack (mainstream AI development choices), strong scalability (easy to add tools/connect to databases).

Application Scenarios: Customer service data analysis (understand ticket distribution/identify problem hotspots), sales lead analysis (customer behavior analysis from CRM data), operation data monitoring (decision support from system logs/user behavior).

Comparison with General RAG:

Dimension	General RAG	Agentic Workflow
Interaction Mode	Single retrieval + generation	Multi-step decision + execution
Controllability	Low	High
Complexity Handling	Simple Q&A	Multi-step complex tasks
Interpretability	Low	High
Scalability	Dependent on vector database	Integrate any API and function

Summary: The project accurately conveys the core value of Agentic Workflow, transforming AI from a passive responder to an active executor. It is an ideal starting point for AI Agent development—with concise code, clear concepts, a modern tech stack, and strong scalability.

Learning Steps: 1. Clone the code and run it to experience; 2. Read the tool definitions and core code for model calls; 3. Try adding new tools; 4. Replace CSV with a real database connection.

Expansion Directions: Visual charts (integrate chart libraries), export functions (Excel/PDF), natural language to SQL, multi-turn dialogue context, permission control (restrict data access by role).