For twenty-five years the internet was built around one assumption: on the other side of the screen sits a human. The browser had tabs because a human opens several things at once, it had themes because a human looks at it, it had extensions because a human likes to customise. Now Cloudflare has released Kitesurf - a browser with all of that stripped out, because the user is no longer a human.

Kitesurf is a browser built for artificial intelligence. It lives in the cloud, has no window, no back button. Instead it worries about things that were never a problem for a person: how much context it can hold, what the tokens cost, how many parallel sessions it can take. The company openly writes that such a browser also has a different security profile - it is vulnerable to attacks through injected instructions, something that literally does not happen to a human browser.

The number worth holding on to is a different one. Cloudflare says the decision to build Kitesurf was made 12 weeks ago. Three months from idea to shipped product, for something meant to replace the layer the entire web passes through. That is not the speed of engineering ambition - it is the speed of a company afraid of being late.

Why Cloudflare and not somebody else

Cloudflare is not a browser maker. Cloudflare is a company through whose network an enormous share of internet traffic passes - it is the doorman, not the tenant. If AI agents start using the web instead of people, then the question of who holds the door becomes far more expensive than the question of who makes a prettier tab icon.

The technical part reveals the same thing. Kitesurf was not written from scratch - it is assembled from other people's parts: a modular rendering engine from Blitz, the CSS parser from Firefox, and Boa JS as an engine written in Rust. Everything else runs on the company's own serverless platform, Workers. Cloudflare also credits Obscura, an open project from which the first prototype was merely a transplant onto their platform. The sales argument is efficiency: less processing power and less memory than Chromium for typical tasks like taking screenshots and extracting HTML. In other words - a cheaper bill for whoever is paying for the agent.

How mature is it? The company boasts that it passes around 215,000 web platform tests and adds hundreds more weekly. It sounds like a lot until you remember that this is a measure of how much is left - every test exists because some page can break in some way. There is a list of pages it renders correctly: Wikipedia, Hacker News, its own blog. That is a start, not a finish.

What remains is the question none of the announcements asks. If browsers are built for machines, who will see the ads? The entire media economy in this region - and not only this one - hangs on the assumption that some human opened the page and looked at it. The agent opens it, grinds it into text, returns it as an answer and never remembers it existed. The question is not technical. The question is who pays when the reader stops being a reader.