How does an AI agent differ from a plain model call?

With a simple call I send input and get an answer. In an agentic system the model can also decide what step to take next: search for data, use a tool, read the result and only then answer. It usually does this in a repeating loop controlled by the application.

For example, a model preparing a text is one call. An assistant that checks product availability in an API, compares the result and asks a follow-up question behaves more like an agent. But the name “agent” doesn't determine the degree of autonomy. I always ask what the model actually chooses and what the code sets in advance.