🤖 Anatomy of an agent

You’ve already talked with an chatbot: you ask one thing, it answers one thing, and the conversation stops there. Each message is a separate round. Useful for getting an answer—but after that, the work is still yours.

One agent it’s something else. You don’t give it a question; you give it a goal — and it doesn’t need to be long. Three well-written sentences are worth more than a vague page.

"Thinking" here is just an easy way to put it. Under the hood, the agent repeats a very simple cycle, with three steps that revolve around: observe the situation, decide the next step and act. Then it looks at the result and starts again.

Each cycle brings you a little closer to the goal. It's the same thing you do when assembling a piece of furniture: look at the part, decide where it fits, screw it in, check it—and move on to the next one. The agent does this very quickly and without getting tired.

On its own, an AI only speaks: it produces text. For it to do something in the world, you need tools — concrete actions someone gave it permission to use. Without tools, it’s a brain without hands.

The most common tools are easy to understand: find information on the web, open and edit a file, or run a command on your computer. A tool turns “I’d suggest doing X” into “X was done.”

Here’s the part that most surprises beginners: when one step goes wrong, the agent notices and tries another way — without you asking. It doesn’t give up at the first stumble; it returns to the cycle, observes what failed, and adjusts.

That’s exactly what makes it feel like it’s “really working.” Just like an employee: if the first door is locked, they don’t call you to ask what to do — they try the next one.

The agent is fast and persistent—but not infallible. Two mistakes are common. The first is hallucinate: making up information with complete confidence, as if it were true. The second is stay on the wrong path, repeating the same bad decision with conviction.

That’s why the conductor exists: you’re the one who checks the result. You don’t need to understand how it did it — you need to check whether the result makes sense. Learning to “sniff out” a suspicious result gets you halfway there.

"But isn’t this just automation?" Not exactly. One fixed automation follows a track: it does exactly the steps someone programmed, in the programmed order. If the situation changes, it breaks or does something foolish—because it doesn’t know how to improvise.

One agent adapts. When the situation changes, it notices and chooses another step. Automation is a train on a track; an agent is a driver who steers around a pothole. Each is suited to a different purpose.