A small example
A chatbot can draft a reply to a customer. An agent can read the support ticket, look up the order, check the refund policy, draft a reply, and stop for approval before sending it. The difference is not personality. It is the ability to select and use tools across several steps.
The parts that make it an agent
- A goal or task boundary.
- A model that chooses the next step.
- Tools such as search, a database, a browser, or an API.
- State or memory for the current task.
- A stopping rule and, for risky actions, a human approval gate.
What an agent is not
A long prompt is not automatically an agent. Neither is a chat window with a confident tone. If the system cannot inspect a result, choose the next action, or use a tool, it may still be useful, but it is better described as an assistant or workflow step.
When an agent earns its complexity
Use one when the task repeats, has several clear steps, and can be checked at each risky boundary. Do not start with an agent for a one-off task that a person can finish faster with a checklist. The more freedom the system has, the more important logs, permissions, retries, and approval gates become.