Given a task, an agent can read a situation, call a tool such as a database or an email API, inspect the result, and decide what to do next, including asking a person. That flexibility is the point and the risk.
Most businesses that think they need an agent need a fixed automation. An agent is right for a narrow set of multi step problems, and only when built with typed tools, limits on what it can touch, evaluation cases and a human approval point for anything hard to reverse.
When an agent is right, it is right with a narrow scope, and it is measured on the same tasks every week so drift is caught early.