How AI Agents Actually Work

“Agent” sounds like magic. It is not.

It is a normal AI call, the same kind that powers ChatGPT, running in a loop. Let me show you what is actually under the hood.

One AI call

You type a message. It gets sent to an AI model like OpenAI or Claude. The model replies.

flowchart TD
    A[You type a message] --> B[It is sent to the AI model]
    B --> C[The model replies]

The interesting part is what you put in that message. Along with your request, you hand the model a goal and the tools it is allowed to use. That whole thing is just a prompt, and this is what you actually send to the server (simplified):

Example prompt (what you send)

{
  "goal": "Book me a table for two on Friday",
  "tools": ["check_calendar", "search_restaurants", "send_email"],
  "instruction": "Reply with an answer, or the tool to run and its inputs."
}

The server replies, also as JSON, with what it wants to do next:

The reply (what comes back)

{
  "run_tool": "search_restaurants",
  "inputs": { "day": "Friday", "people": 2 }
}

Notice the model did not actually run anything. It just said which tool to use.

The loop

That is the cue for the loop. A piece of plain, deterministic automation reads the reply, actually runs search_restaurants, and drops the result into the next message back to the model.

flowchart TD
    A[Model says: run this tool] --> B[Automation runs the tool]
    B --> C[Result is added to the next message]
    C --> D[Model reads it and picks the next step]
    D --> A

The model reads the result, decides what to do next, maybe “run send_email,” and the loop continues until the goal is met.

It is just ChatGPT with hands

That is the whole trick. It is no different from the ChatGPT window you already use: you type, it replies.

The only thing added is a deterministic wrapper. Whenever the model says “run this tool,” the wrapper runs it and feeds the result back.

The model is the only unpredictable part. It decides what the next message should do. Everything it asks to run is executed by plain, predictable code.

So an agent is not a new kind of intelligence. It is the AI you already know, given a loop and a set of hands.