What is an agent

The word "agent" gets thrown around a lot. Strip it down: an agent is a language model that has been given (1) a goal, (2) a set of tools it can call, and (3) a loop that lets it call those tools more than once before giving you an answer. That's it. Everything else is engineering on top of those three things.

A chat box has one of those things: it has a model. You give it a prompt, it gives you a string back, and the loop ends. If you want it to read your calendar, you have to paste your calendar in. If you want it to send an email, you copy the draft and send it yourself. The model is the engine, but you are the chassis.

Three pieces, in order

• Goal — a clear instruction about what the loop is trying to achieve, written in the system prompt. "Triage incoming support tickets and route the simple ones to the right canned reply."
• Tools — discrete functions the model can call by name with structured arguments. read_inbox(), search_kb(query), send_reply(ticket_id, body). The model doesn't run these; your code does, and hands the result back as a message.
• Loop — the model proposes a tool call, your code runs it, the result goes back to the model, and it decides whether it's done or whether it needs to call another tool. Repeat until the goal is met or the loop budget runs out.

Why "loop" is the word that matters

The single most useful mental shift from chat → agent is the loop. A chat is one-shot. An agent is N-shot, where N is decided dynamically by the model. That means an agent can read, then decide to read more, then write, then check its own work, then ship. A chat would need a human to drive each of those steps.

It's also where most of the failures live. Loops can wander, repeat themselves, or get stuck calling the same tool over and over. The discipline of building agents is mostly the discipline of writing tight loops.

What this means for week one

When you write your brief in the next lesson, you'll pick a task in your week and describe it in those three pieces: what's the goal, what tools would the agent need, and what does "done" look like at the end of the loop. If you can't say all three in a paragraph, the task probably isn't right for an agent yet — it's right for a script, or for you.

Don't worry about model choice, framework choice, or any of that. Those are decisions you make later, after you know what you're building. Week one is about teaching your eye to spot agent-shaped problems.