What is an AI agent?

You ask a chatbot where to eat tonight. Usually you get a list. But some assistants now offer something else: they'll check which places are open, compare two menus, and book the table for you. That small offer marks a real shift. You've stopped asking for an answer and started asking for something to be done. Software built to do that is called an artificial intelligence (AI) agent.

Most definitions come down to three parts.

• A goal: something to get done, like booking a table for four at seven.
• Steps: the agent breaks the goal into smaller moves and decides their order.
• Tools: the programs it can use along the way, such as search, a calendar, or a booking site.

Google's machine learning glossary puts it in one line. An agent is software that can reason about what you ask, then plan and carry out actions on your behalf.¹³

What makes it an agent is a loop. It tries a step, looks at what happened, and decides what to do next. If the first restaurant is full, it doesn't stop and report failure. It tries the second one. Anthropic, the company that makes the Claude models, describes this in its guide for builders: at each step, an agent checks real results from its tools to judge how it's doing. That loop is also why an agent takes longer than a chat reply. It's doing several things, one after another.²

On its own, a language model can only write text. Tools let it reach anything else. A search tool lets it look things up, and a file tool lets it open a document. Others let it send an email, add a calendar event, or click through a website. Each tool is a door that someone chose to open, so an agent can only do what its tools allow.

Not every useful agent is clever. Many follow a short, fixed path that a person designed: read the email, pull out the date, add it to the calendar. Anthropic's guide calls those workflows, and saves the word agent for systems that choose their own steps. Its advice to builders is to find the simplest solution that works, and add complexity only when it's needed. So a product that calls itself an agent may be a sensible set of fixed steps, and that's often a good thing.²

With a chatbot, the main question is whether the answer is right. With an agent, two more questions join it.

• What did it do? A good agent shows its steps, so you can see which sites it visited and what it changed.
• What was it allowed to touch? An agent that can only read your calendar is very different from one that can also send email in your name.

You can often set those limits yourself. And it's reasonable to expect a pause for your approval before a step you can't undo, like paying for something or sending a message.²

You may already have used one. Research modes that search dozens of pages and write up what they found are agents. So are coding helpers that edit files and run tests, and booking or shopping assistants that fill in forms for you. Names and features change month to month, so it helps to describe them by what they do. The next time an assistant offers to take care of something for you, ask it those two questions before you say yes.¹