Cloudflare Built A Browser That Throws Away The Pixels

Cloudflare built Kitesurf, a browser for AI agents that treats drawing the page as optional, and the honest reaction is that this is common sense. An agent operating a website does not need to know what the website looks like. What it needs is the structure underneath, and if some of your information only ever existed in the painted pixels, this is the moment it stops existing at all.

Kitesurf parses the page but skips what agents never use

Kitesurf launched on August 6, free during beta through Cloudflare's Browser Run. It is a real browser in the sense that it parses HTML, runs JavaScript, and applies CSS, and it passes over 215,000 web-platform tests. What it is not is Chromium: it runs on Cloudflare's own Workers infrastructure rather than a packaged browser engine.

What it leaves out is the interesting part. In Cloudflare's words: "AI doesn't care about tabs, themes, browser extensions, or synchronization across devices." And more directly: "Structured, machine-readable content is important, but visual perfection, smooth 60-fps scrolling is not."

The tradeoff they measured says something about priorities. Against Chromium, at the median, Kitesurf uses 3.1 to 7.0 times less processing power and 4.7 to 7.0 times less memory, and takes 1.7 to 1.8 times longer to finish. They accepted a slower browser to get a cheaper one. That is a sane trade for a visitor with no eyes and no patience to lose, and it tells you what the people building agent infrastructure think matters.

The visual layer was always one interface, and it often drifted

"Agents do not need the visual layer" gets heard as a claim about design, and it is not one.

The visual layer was one interface to the structure behind it. It was one rendering, made for one kind of visitor. The structure was always the thing, and the pixels were a presentation of it.

What went wrong is that the two came apart. The visual layer drifted away from the structure it was supposed to be presenting, until a lot of what a website communicates lived only in the presentation and nowhere else. That drift is why machines struggle with websites, and it has nothing to do with whether agents can see. Products like Kitesurf are confirmation that this was broken all along.

Trust badges are the clearest example

Look at any e-commerce product page and count the trust badges, the guarantee seals, the free-returns reassurance, and the little icons that exist entirely to make somebody comfortable enough to buy.

They are almost always in the visual layer. They are almost never in the product feed, or the schema, or anything else structured.

So a machine reading that page gets a title, a price, some variants, availability, and an image. It does not get the reason a human decided to trust you enough to buy. All of the persuasion is in the paint.

That was survivable when the paint was the only interface anyone used, and it was already costing something before agents existed, because the same information was missing from every other machine-readable surface too.

This is the bill arriving

For some websites, and possibly most, Kitesurf is the bill arriving.

Until now, there was a workaround. An agent driving a real browser could take a screenshot and have a vision model read the picture, which meant visual-only content was hard to reach rather than gone. Building agent browsers that render pages properly so a model could look at them was always the expensive way around a problem that should have been fixed at the source.

Kitesurf removes the workaround by design. That is not an accusation, it is what the product says it does. When the browser an agent uses is built to skip visual fidelity, the content that exists only as visual fidelity is not hard to reach any more. It is absent.

If painted pixels were the only way some important information was communicated, it was always wrong.

Read your schema and your product feed first

The easiest place to start is your schema or your product feed, depending on what kind of website you have. Both, if you have both.

Not because those are the only two things an agent reads. They are not. An agent reads the HTML, the accessibility tree, whatever tools you expose to it, and whatever else it can get. The reason to start with schema and feeds is that they are the places where you already sat down and tried to describe what you are in a form a machine can use. That makes them a mirror. If the important things are thin there, they are almost certainly thin everywhere else a machine looks, and reading them is the fastest way to find that out about yourself.

So read what is actually in there, and ask two questions.

Is there enough information to explain the business, the product, or the service?

And is there enough to convince someone to take the action you want them to take?

That second question is the one people skip. Machine-readability gets treated as a completeness problem, a checklist of fields populated. But an agent recommending a product to the person who asked it for one is making a case, and it can only make that case out of what it found. If everything persuasive about your offer lives in the theme, the agent has your specifications and none of your argument.

The more advanced version of this check is reading the raw HTML and the accessibility tree, which is where you find out what a machine actually receives rather than what your CMS believes it published. That is worth doing. But schema and feeds are where most people should start, because they are legible without tooling and the gaps are obvious once you look.

Look at the four weeks before Kitesurf launched. On July 9, OpenAI retired ChatGPT Atlas, the standalone AI browser it had launched nine months earlier, which I wrote about at the time as the clearest sign that AI browsers were backward because agents never needed the visual layer. Atlas was a browser built so a person could watch an agent work. Kitesurf is a browser built so nobody watches. One of those died and the other launched inside a month, and neither event needed my opinion to make the point.

What Kitesurf adds is the price. Running agents through a full browser was expensive, and Cloudflare cut that cost to a fraction by dropping the part agents were never using. Cheap infrastructure is what turns a capability into a habit.

So the question stops being whether machines will read your website without looking at it. They will, and it will cost them almost nothing. The question is what they find when they do.