3.5 What Is an AI Agent?

An ordinary chat assistant often receives a question and returns a response. An AI agent works toward a goal by planning steps, using permitted tools, observing results, and deciding what to do next.

Think of ordinary chat as an assistant that offers advice and an agent as an assistant with a task list and access to tools such as a calendar or search. Arranging a team dinner might involve gathering requirements, searching restaurants, comparing options, producing a shortlist, and booking only after approval.

Tools, permissions, and approval

The word “agent” does not grant access to everything. Its actions depend on connected tools and permissions. Reading a calendar differs from editing it; drafting an email differs from sending it.

The safest principle is least privilege: grant only the access needed for the current task. Do not allow editing when reading is enough, or automatic sending when a draft is enough. Payments, deletion, public publication, outgoing messages, and changes to important records should have clear human approval.

An agent can misunderstand a goal, choose the wrong tool, or amplify an early mistake over several steps. Requiring a visible plan, limiting its action scope, keeping an activity record, and pausing at important checkpoints is safer than saying, “Handle everything for me.”

Not every chat product is an agent, and not every automated workflow uses a large model. Ask whether the system chooses next steps toward a goal and calls tools that affect external objects.