Beyond the label
The word agent is used for many different products. A useful working definition is a system that can choose actions toward a goal, observe what happened, and decide what to do next. The amount of independence can vary considerably.
Tools turn text into action
A language model can propose a search query. A tool lets the surrounding system run that search and return the result. Other tools might read a file, calculate a value, or prepare an edit. The application decides which tools exist and what they are allowed to touch.
The loop needs limits
An agent can repeat a cycle of planning, acting, and checking. That flexibility is useful, but a loop also needs a budget and a stopping condition. Without them, a small misunderstanding can become a long series of unhelpful actions.
Permission is part of the design
Reading a public page and publishing a message are different kinds of action. A well-designed system makes those boundaries explicit. It should be clear which steps happen automatically, which require review, and how to undo a change where that is possible.
The takeaway
Ask less about whether a product calls itself an agent and more about what it can do, what it can access, and how it knows when to stop. Those details tell you much more than the label.
