The core idea

An agent is an LLM that can act, not just answer.

A plain model returns text. An agent wraps the model in a loop that lets it take actions in the world, observe what happened, and decide what to do next.