A useful definition

The word "agent" gets thrown around loosely, so here's a definition that actually pins it down: an agent is an LLM placed in a loop with access to tools. A plain chatbot answers your question once and stops. An agent is given a goal and the freedom to take steps toward it — it can plan, call a tool (search the web, run code, hit an API), look at what came back, and decide what to do next, repeating until the goal is met or it gives up.

That loop is the whole difference. The model isn't just producing text; it's deciding what to do and then doing it.

From answering to acting

Consider "What's the weather in Paris, and should I pack a coat?" A chatbot answers from memory, which might be out of date. An agent recognizes it needs live data, calls a weather tool, reads the result, reasons about the temperature, and answers — grounded in a real lookup it performed itself.

Chatbot = one question, one answer. Agent = a goal, and the freedom to take steps toward it.

Power and peril, same coin

This shift from answering to acting is exactly what makes agents useful and what makes them dangerous. An agent that can book travel can also book the wrong travel; one that can run code can run harmful code; one that reads web pages can be manipulated by what it reads. Everything later in this course — guardrails, limits, evaluation — exists because acting in the world has consequences that answering a question doesn't.

What agents are good for

The sweet spot is tasks that need several steps, external information, or tools — research across sources, multi-step automation, "look this up, then do that." For a single-shot answer, a plain prompt is simpler and safer; reach for an agent when the task genuinely requires doing, not just saying.

An agent is a model with hands. That's the source of both its usefulness and every risk you'll spend the rest of this course managing.

Try this: Take a request you'd normally type into a chatbot and ask: does answering it require looking something up or taking an action? If yes, it's an agent-shaped task; if it's pure knowledge, it isn't. That one question tells you when an agent is the right tool.