What is an AI agent?

A plain question-and-answer exchange is one round: the model reads a prompt and writes a response. An agent is the same kind of model wired into a loop instead: it can take an action, like calling a tool to search or read a page, look at what came back, decide what to do next, and repeat, often several times, before it decides the task is finished and gives a final answer.

It's worth being clear that an agent isn't a smarter or different model underneath, it's a way of running an ordinary model repeatedly with tools attached, so it can still misread a page or state something wrong partway through a multi-step task (see hallucination). For a site owner, agents matter because a shopping or research agent may visit and read a page directly during its task, rather than only seeing a summary from a search tool, which means the ordinary readability of that page, whether the content actually renders and loads properly, is back in play (see thin-content, server-side-rendering).