SITUATION

In our previous section, we demonstrated how to use ChatGPT API to deliver app review analysis by summarising each review from app users (link) and how to use OpenAI Assistants API to develop a very simple chatbot to answers questions from an uploaded document (link).

However, real products often need more than answers— they need actions. For this two-part blogs, we will build a simple agent that can track job applications.

• Part 1 - this post covers the fundamentals: what an agent is, why we insist on structured JSON outputs, and how function calling lets the model fetch real data instead of guessing.
• Part 2 applies those techniques to the job-application tracker end-to-end.

If you only take one idea from Part 1, make it this:

Reliable agents = (Structured JSON) + (Function Calling) + (Short, controlled loops).

It is recommended that one has a basic Python coding experience and access on OpenAI API keys. Please refer to this blog about how to setup OpenAI API.

What is an agent? Why not just prompt the API directly?

Direct prompting - the “chat” way

• Send a prompt like “Create a job application at Apple” or “What’s the weather in Paris?”
• And the model replies with free-form prose

This is perfect for brainstorming and summaries. But it is brittle when we need to extract fields, call APIs, or write data. Free-form text is unpredictable: formatting drifts, important fields go missing, and the model might confidently “know” facts that are wrong or outdated.

An agent

An agent treats the model as a planner that:

1. returns a structured JSON object (the contract),
2. requests tool calls when it needs real data (APIs we control),
3. works in short steps: plan → call tool → read the result → finish.

User text
  → LLM interprets the request
  → returns JSON you defined (contract)
  → calls your tool(s) for real data when needed
  → returns a final structured result your app can store or act on

Benefits of using agent instead