User Agent

When you open a webpage, you don’t fetch it yourself — your browser does that for you. It sends a request to the server, receives the response, and displays it to you. This program, which acts on your behalf, is called a user agent. The word “agent” here simply means: someone or something that acts on your behalf. In the classic web, that’s almost always the browser. In the modern AI world, the term has acquired a second, related meaning: there, AI systems are called user agents when they independently carry out multiple steps to complete a task — for example, booking a trip or sorting an email inbox.

Why servers want to know who’s asking

Every request that a user agent sends to a server contains a short text stamp — also called a “user agent.” It states which program is making the request: for example, Chrome on Windows or Safari on an iPhone. Servers use this information to deliver the appropriate version of a page. A mobile version of a website is delivered because the server recognizes the user agent string and knows: this is a mobile browser.

That sounds harmless, but it has consequences. Advertising companies use the user agent string to detect device types and track more precisely. Privacy tools like the Tor Browser therefore deliberately disguise their user agent string, making users harder to distinguish from one another. Search engine crawlers — i.e., programs that automatically scan webpages — identify themselves with their own user agent string so that website operators can distinguish them from regular users.

From browser to independently acting AI system

In AI research, “user agent” today means something more sophisticated. An AI system is considered a user agent if it doesn’t just answer a single question, but instead plans and executes a chain of actions on its own. It opens webpages, fills out forms, evaluates results, and makes intermediate decisions — all without the user confirming every step. The human only specifies the goal.

The difference from a regular chatbot is substantial. A chatbot answers a question. An AI user agent carries out a plan consisting of many individual steps. In doing so, it uses tools: web browsers, calendar apps, code execution environments. Researchers also speak of “agentic behavior” in this context — the model behaves like an actor, not like a reference work.

An important question here is how much control the human retains. The more steps a user agent carries out independently, the more can go wrong without anyone intervening. That’s why developers are intensely discussing so-called “guardrails” — boundaries that define what a system is allowed to decide on its own and when it must ask for permission.

User agents in products and headlines

In everyday life, you encounter the classic user agent mainly in your browser’s developer tools. There you can view the user agent string of your browser and even change it manually — websites will then see you, for example, as a mobile user even though you’re sitting at a laptop.

In tech news, the term increasingly appears in an AI context. Google's “Project Mariner,” Anthropic’s “Computer Use,” and OpenAI’s Operator features are examples of AI user agents that remotely control a browser and are meant to shop or research online on the user’s behalf. Apple has announced plans to integrate similar features into its operating system so that Siri can handle complex tasks across multiple apps.

Both meanings of the term are, incidentally, directly connected: the AI user agent internally uses a real browser user agent to retrieve webpages. The concept of acting on someone’s behalf is the same in both cases — only the complexity of the task has grown.