Building the Browser Layer for AI Agents

The web was built for people.

Websites are designed around screens, buttons, menus, forms, pop-ups, and visual layouts that make sense when a person is sitting in front of a browser. For decades, that was a safe assumption. Software talked to software through APIs. People used the web.

AI agents are beginning to blur that distinction.

An agent asked to book a flight, research a company, submit a form, manage an account, or buy something online eventually encounters the same problem: much of the information and functionality it needs is trapped behind a website designed for a human being.

APIs solve part of the problem, but most of the web doesn’t have one. Even when an API exists, it rarely exposes everything a person can do through the browser.

Browser Use is building the infrastructure that lets agents operate the web anyway.

A Weekend Project Takes Off

Magnus Müller and Gregor Žunič met while studying data science at ETH Zurich. Both had recently left previous startups and found themselves at ETH’s Student Project House, an incubator where students are encouraged to work on ambitious projects rather than coursework. They started talking about what would happen as AI agents moved beyond answering questions and began trying to do things on the web.

Magnus had been building bots and web automations since he learned to code and describes himself as a scraping hobbyist and machine learning researcher. Gregor came from physics before pursuing his master’s in data science at ETH. Together, they started experimenting with a basic question: could they give language models an interface that made the web easier to understand and control?

They built the first prototype in a matter of days and put it on Hacker News.

It took off.

What started as a side project quickly became one of the largest open-source projects for browser agents. The response told Magnus and Gregor something important: developers didn’t just need traditional browser automation. They needed infrastructure that could help an intelligent system understand an unfamiliar website, decide what to do, and act on it.

Giving Agents a Browser

Browser Use approaches the web differently from traditional automation tools.

Rather than asking a model to interpret every webpage primarily through screenshots, Browser Use can turn the structure of a website into a text-like representation that an agent can reason about. The agent can understand what is on the page, decide what to do, take an action, observe what happened, and continue.

That matters because the browser is an unusually powerful interface. A company may never build an API for its scheduling system, supplier portal, insurance dashboard, or internal application. But if a person can log in and use it through a browser, an agent potentially can too.

Traditional browser automation works well when the path is known in advance: go to this URL, find this field, click this button. Agents are being asked to handle tasks where neither the path nor every step can be specified beforehand.

They have to reason their way through the web.

Browser Use sits between that reasoning and the browser itself.

A Team Built Across the Web and AI

Magnus’s background makes the company feel particularly natural. He spent years building bots, scraping tools, and web automations, worked in R&D around process automation, and studied machine learning and data science. He understood firsthand how difficult it is to make software reliably interact with websites.

Gregor brought a complementary way of thinking about the problem. He studied physics before moving into data science, a combination that trained him to break complicated systems into something more structured and tractable. As CTO, he has become deeply focused on the technical problems required to make browser agents faster, more reliable, and capable of operating at scale.

But perhaps the most telling part of the team story is how Browser Use itself began. Magnus and Gregor didn’t spend months designing a platform around a hypothetical market. They built a small experiment, released it to developers, and watched people start using it.

That developer-first instinct has shaped the company ever since.

The Web as an API

The larger opportunity becomes clearer when you consider how much of the world’s software is accessible through a browser.

Millions of websites and applications contain functionality that has never been exposed programmatically. Humans bridge those systems every day by reading information from one screen, typing it into another, downloading a file, uploading it somewhere else, or clicking through a workflow.

For software, those interfaces have historically been walls.

Agents turn them into potential endpoints.

If an agent can reliably operate a browser, a website begins to function a little like an API, even when nobody built one. That dramatically expands the universe of tasks software can perform.

It also makes the browser an increasingly important piece of AI infrastructure. Models provide intelligence. Other systems provide memory, identity, payments, and communication. But when an agent needs to interact with the enormous portion of the world that still lives behind a graphical interface, it needs a browser it can control.

Browser Use is increasingly building the infrastructure around that interaction as well, including the systems required to run browser agents reliably and at scale.

Why We Invested

We invested in Browser Use because we think the ability to act on the web will become a basic primitive for AI agents.

Magnus and Gregor arrived at the opportunity with unusually relevant backgrounds, but what really stood out was how quickly developers validated the need. They built something in a matter of days, put it into the world, and watched an open-source community form around it.

There are many possible futures for agents, and it’s still early enough that nobody knows exactly what the dominant applications will look like. But almost all of those futures share one requirement: agents need to interact with the world as it exists today, not just the parts rebuilt specifically for AI.

A huge portion of that world lives in the browser. Browser Use is building the layer that lets agents in.

Visit Browser Use at https://browser-use.com/.