What is an AI "agent"?

A chatbot usually responds to one request and waits for you. An AI agent goes further: it can choose and use tools, keep track of what happened, and repeat steps toward a goal.

A simple agent loop looks like this:

1. Read the goal and the current state.
2. Choose a next action, such as searching approved documents or updating a file.
3. Use a tool and observe the real result.
4. Compare that result with a stopping condition.
5. Continue, ask a person for help, or stop.

The loop is only part of the system. The available tools, permissions, memory, environment, stopping conditions, and human review determine what the agent can actually do. Some products also call fixed multi-step workflows “agents,” so the label is less important than understanding which actions the system may take.

An agent should have boundaries: limited permissions, a retry or cost limit, and a clear point where a person must approve a consequential action. Its own claim that the job is finished is not proof; the result still needs to be read back from the real target.

For a broader starting point, see Learning AI for Everyday Life. Anthropic’s Building effective agents is a useful technical source for the distinction between predefined workflows and agents that dynamically direct their own process.

Accuracy note: reviewed September 23, 2026. Agent terminology and product capabilities change quickly.