NEWS · AUGUST 10, 2026 · ARTIFICIAL INTELLIGENCE

Cloudflare shipped a browser written for agents: Kitesurf

On August 6, 2026 Cloudflare announced Kitesurf, a browser designed for AI agents rather than people and running on Workers. There are no tabs, themes or extensions. In their place there is low CPU, low memory and machine readable content. It is free to use during the beta.

01 · WHAT HAPPENED?

A browser for the model, not the human

Cloudflare introduced Kitesurf in a blog post dated August 6, 2026. The title also states where the browser runs: on Cloudflare Workers, inside V8 isolates. The company describes Kitesurf as a stateless, highly scalable and cost effective browser.

The design choice is summed up in a single line. According to Cloudflare, AI does not care about tabs, themes, browser extensions or synchronization across devices. What it cares about is token count, context windows, scalability, performance and cost. The company adds that visual perfection and smooth 60 fps scrolling are not required, while structured, machine readable content is.

TechCrunch reported on August 7, 2026 that Kitesurf was built in 12 weeks on top of the company serverless Workers platform. Per that report the browser renders TodoMVC, Wikipedia, Hacker News, the Cloudflare Blog and much of the Cloudflare dashboard.

02 · THE DETAILS

Three parts, two Rust engines and one benchmark table

The architecture splits into three main components. Engine handles Chrome DevTools Protocol communication and session state. PageScript spins up isolated JavaScript execution environments with Dynamic Workers and processes page logic through the Rust based Blitz library and Stylo, the CSS parser from Firefox. PageRenderer turns the computed page object into pixels. A fourth component, SandboxOutbound, sits on the security side controlling outbound requests and cookies. JavaScript evaluation uses Boa JS, an ECMAScript engine written in Rust.

Cloudflare published a benchmark comparing Kitesurf against a warm pool Chromium. On screenshots Kitesurf spends 380 ms of CPU against 1,173 ms for Chromium, a 3.1x gap. On HTML extraction it is 229 ms against 877 ms, a 3.8x gap. The memory gap is wider: 57.8 MiB against 271.0 MiB on screenshots, which is 4.7x, and 39.4 MiB against 273.7 MiB on HTML extraction, which is 7.0x.

There is a column where Kitesurf comes out behind. On wall clock time a screenshot takes 1,148 ms against 637 ms on Chromium, and HTML extraction takes 820 ms against 472 ms. That is roughly 1.7 to 1.8 times slower per operation. The source does not hide that gap, it puts it in the same table.

On compatibility, Kitesurf passes around 215,000+ Web Platform Tests with hundreds more added weekly, and coverage is strongest in CSS, DOM, HTML, SVG and XMLHttpRequest. Existing Puppeteer, Playwright and chrome-remote-interface setups connect by adding a browser=kitesurf parameter to the endpoint. It is free through Browser Run during the beta with per account limits. The gaps are stated plainly too: video playback, WebGL rendering, TLS fingerprint based bot challenges and long lived authenticated sessions that need persistent state are not supported yet.

03 · WHY IT MATTERS

Two different readers now visit your site

The real story here is not a product but an admission. An infrastructure company writing a browser for agents from scratch, rather than adapting the human one, says this traffic is not being treated as a passing wave. That assessment is ours, the sources make no such comment.

Second, the unsupported list is effectively a warning list. If WebGL is not rendered and video is not played, a site that leans mostly on those layers stays largely empty in an agent view. The source does not write this as site advice, the inference is ours.

Third, cost. A multiple fold drop in CPU and memory means browsing a page gets cheaper on the agent side. Anything that gets cheaper gets more common. A growing share of automated clients among your daily visitors is therefore a reasonable expectation.

Fourth, security. The TechCrunch report names prompt injection among the security concerns specific to agent browsers. For a client that reads the text on a page and then acts on it, instructions planted in that page are a new attack surface.

04 · TURKEY

What it means for businesses in Türkiye

The assessment below is not in the sources, it is our reading. The sources note no Türkiye specific breakdown.

Two patterns we see often on business sites here intersect directly with this news. First, most of the content loading later through JavaScript. Second, contact and service details being embedded inside images rather than text. Neither causes trouble for a human visitor, but both can make that information invisible to a client looking for machine readable content.

We suggest three practical checks. First, look at what remains on your page with JavaScript disabled, since what remains is close to the minimum an agent sees. Second, make sure service descriptions, address, opening hours and contact details exist as plain text on the page. Third, check whether blunt server side rules that block automated clients are also cutting off the agents you actually want.

One measurement note to close. Agent traffic is not separated from human traffic in most analytics setups. In a report that cannot make that distinction, conversion rate can look like it is falling for no reason. That line is our comment as well.

The UNALSOFT take

In our agentic AI work we have been telling clients the same thing for a while: a site is now written for two readers at once, one human and one machine. Kitesurf gives a technical description of that second reader. What it cares about, and what it never reads, are now written down in front of us. Our order of work follows from that: content that also exists server side, a structured data layer set up correctly, critical information not trapped inside images, and blocking rules chosen deliberately. None of this looks like design work, yet these are exactly the things that decide how much of your site gets read on the machine side. The related work runs alongside web design.

How much of your site does a machine actually read?

Let us check together how your content looks to an agent.

Message on WhatsApp