A browser Cloudflare built with AI agents in mind, not people. It uses much less CPU and memory than Chromium and takes a bit longer in wall-clock time.

Why would anyone want a slower browser?

Because an agent doesn't get impatient. If you run a thousand sessions in parallel, what runs out is memory, not patience, so it pays to save memory even if each session takes a bit longer.

How much slower is it, exactly?

1.7 times when extracting HTML and 1.8 times when taking a screenshot, according to Cloudflare's benchmarks on 14 sites.

How do I know which parts of my product are dead weight if an agent uses it?

With a four-column audit: for each feature, whether it exists because a person is waiting, whether it's unnecessary or in the way when an agent consumes it, and how much it costs to keep. Whatever lands on “yes and unnecessary” is human-interface debt.

Do I have to delete what's unnecessary?

Not necessarily. You have to know how much it weighs, because that cost shows up late and without warning, for example when someone asks why the API costs what it costs.

August 11, 20263 min read

Cloudflare built a browser 1.7 times slower, on purpose

Kitesurf is a Cloudflare browser built for AI agents that takes 1.7 to 1.8 times longer than a regular one, on purpose, because its user doesn't get impatient and what matters is spending less CPU and memory.

Almost everything we discuss in product assumes there's a person on the other side, waiting.

The spinner, the transition animation, the three-line summary, the pagination. None of it solves a problem: it solves the impatience of someone looking at a screen.

Cloudflare just published a browser that throws all of that out and comes out slower on purpose, because its user is an agent and an agent doesn't get impatient.

The numbers

It's called Kitesurf. Cloudflare measured it against Chromium on a set of 14 sites.

3.1xless CPU to take a screenshot380 ms vs. 1,173 ms
3.8xless CPU to extract the HTML229 ms vs. 877 ms
4.7x to 7xless memory than Chromium57.8 vs. 271.0 MiB, and 39.4 vs. 273.7 MiB
1.7x to 1.8xslower in wall-clock time1.7x extracting HTML, 1.8x on screenshots

It took them 12 weeks to build. The first commit is from May.

What they removed says more than what they optimized

Tabs. Themes. Extensions. Sync. Pixel-perfect CSS. 60 fps scrolling.

All of that exists for one reason: someone was waiting on the other side. An agent doesn't get impatient, doesn't watch, and doesn't switch to another tab to see if it loads.

The equation, flipped

Same product, different user, opposite decision.

One person

Time is expensive. Memory is free. It makes sense to spend whatever it takes so the page shows up right now.

A thousand parallel sessions

Time is free. Memory is everything. It makes sense to take a bit longer if that lets you run more sessions together.

It's the cleanest product decision I've seen this year, and it isn't technical. It's asking who the user really is, and living with the answer.

The uncomfortable part

Part of your product doesn't solve a problem. It solves a person's wait: spinners, transition animations, three-line summaries, pagination, autocomplete.

Human-surface audit

A table with one row per feature and four columns. It takes an afternoon.

The four columns of the audit
Column	What you answer
1. Feature	The name of what you're reviewing.
2. Human wait	Does it exist because a person is waiting? Yes or no.
3. If an agent consumed it	Is it unnecessary, in the way, or still useful?
4. Cost of keeping it	Rendering, tokens, latency and maintenance.

Everything that lands on “yes + unnecessary” is your human-interface debt. You don't have to delete it: you have to know how much it weighs before someone asks why your API costs what it costs.

The hard part for Cloudflare wasn't the rendering engine. It was accepting that the user had changed.

Frequently asked questions

What is Kitesurf?: A browser Cloudflare built with AI agents in mind, not people. It uses much less CPU and memory than Chromium and takes a bit longer in wall-clock time.
Why would anyone want a slower browser?: Because an agent doesn't get impatient. If you run a thousand sessions in parallel, what runs out is memory, not patience, so it pays to save memory even if each session takes a bit longer.
How much slower is it, exactly?: 1.7 times when extracting HTML and 1.8 times when taking a screenshot, according to Cloudflare's benchmarks on 14 sites.
How do I know which parts of my product are dead weight if an agent uses it?: With a four-column audit: for each feature, whether it exists because a person is waiting, whether it's unnecessary or in the way when an agent consumes it, and how much it costs to keep. Whatever lands on “yes and unnecessary” is human-interface debt.
Do I have to delete what's unnecessary?: Not necessarily. You have to know how much it weighs, because that cost shows up late and without warning, for example when someone asks why the API costs what it costs.

Sources

Cloudflare: KitesurfOfficial post with the full benchmarks, the 14 sites and the breakdown by task.