Agent

An agent is a loop in which a model can choose the next action, inspect the result, and continue toward an objective.

Tiny example

request → model chooses search → search result
        → model chooses lookup → customer record
        → model drafts answer → stop

The surrounding application owns the loop, state, tool execution, permissions, retries, and stopping conditions.

What it is not

An agent is not automatically autonomous, reliable, or appropriate for every workflow. A fixed sequence is usually easier to test when the required steps are already known.

FDE note

Bound the number of steps, expose progress, require approval for consequential actions, and make side effects safe to retry.