A language model alone produces text. An agent places that model inside an application with tools and a loop.

The model proposes a tool request. The application checks permissions and runs the tool. Its result returns to the context, and the model chooses what to do next.

Retrieval can add current or specialised documents without changing the model's weights. It can run once before an answer or appear as one tool inside a longer agent loop.

Each step can help the next one. Each mistake can also steer the next one. The following chapter asks which tasks suit that loop and where human control belongs.