Agent

An agent runs a loop: the model chooses an action, a tool executes it, the result comes back into the context-window, and the loop repeats until the goal is met or a limit is hit. The tools are ordinary functions - search, read a file, run a query, call an API.

What you gain is the ability to hand over a task without specifying the steps. What you take on is that errors compound. A wrong step at position three poisons everything after it, and the model will usually keep going confidently.

Production agents therefore need boundaries in the prompt and in the harness: a step limit, a definition of done, explicit handling for a failed tool call, and a human checkpoint before anything irreversible.