What Makes Something "Agentic"

Every course on this site up to now has treated a model's output as something a person reads. Zero to GPT built the actual mechanism, a forward pass through the network produces a probability distribution over the next token, and sampling picks one, token by token, until generation stops. Nothing about that mechanism cares what the tokens mean once they're picked. An agent is what you get when some of those tokens get treated as an action instead of a message, text a program parses and does something with, rather than words a person reads.

Without anything to call, the model did what a careful model should, it said plainly that it has no real-time access to the weather, and suggested a few places a person could look instead. Give the exact same underlying question a tool to reach for, get_weather(city), and tell it to respond with only a JSON call when it needs live information, and the model's output stops being prose entirely, it's a single line of structured JSON, nothing else.

Nothing about the model changed between those two calls, same weights, same next-token mechanism. What changed is what it was told it could reach for. This is the whole idea this course builds on, an agent isn't a different kind of model, it's the same generation process, pointed at a world where some of its own output gets executed instead of just displayed.