Agent

An agent runs a loop. The model picks an action, a tool carries it out, the result lands back in the context-window, and it goes round again until the goal is met or a limit is hit. The tools are plain functions: search, read a file, run a query, call an API.

You gain the ability to hand over a task without listing the steps. You also accept that errors pile up. A wrong step at position three poisons everything after it, and the model usually carries on with confidence.

So a production agent needs limits in the prompt and in the code around it: a step cap, a definition of done, a plan for a failed tool call, and a person who signs off before anything that cannot be undone.