Days after the fake story about Cursor building a web browser from scratch with GPT-5.2 was debunked. Disbelief should be the default reaction to stories like this.
The end result: Me and one agent (codex) managed to build something more or less the same as Cursor's "hundreds of agents" running for weeks and producing millions of lines of code, in just 20K LOC (this includes X11, macOS and Windows support). Has --headless, --screenshot, handles scaling, link clicking and scrolling, and can render basic websites mostly fine (like HN) and most others not so fine. Also included CI builds and automatic releases because why not.
The repository itself is here and should run out of the box on most modern OSes, downloads can be found at the Releases page: https://github.com/embedding-shapes/one-agent-one-browser
Here's a screenshot I took with it: https://bsky.app/profile/simonwillison.net/post/3mdg2oo6bms2...
one-agent-one-browser-Linux-X64 1.14 MB
one-agent-one-browser-macOS-ARM64 1.02 MB
one-agent-one-browser-Windows-X64.exe 847 KB
I’m no expert but it seems like a strange choice to me - using a mutex around an MPSC receiver, so whoever locks first gets to block until they get a message.
Is that not introducing unnecessary contention? It wouldn’t be that hard to just retain a sender for each worker and just round robin them
A poc that would usually take a team of engineers weeks to make because of lack of cross disciplinary skills can now be done by one at the cost of long term tech debt because of lack of cross disciplinary knowledge.
This is where I wish we spent more energy, figuring out better ways to work with the AI, rather than trying replace some parts wholesale with AI. Wrote a bunch more specifically about that, while I was watching the agent work on the browser itself, here: https://emsh.cat/good-taste/ (it's like a companion-piece I guess)
Although the tell is obvious if you spent one second looking at https://github.com/nkuntz1934/matrix-workers. That misaligned ASCII diagram, damn.
Why is Cloudflare paying this guy again, just to vibe a bunch of garbage without even checking above the fold content in the README?
It's getting outright frustrating to deal with this.
Fine, random hype-men gets hyped about stuff and tweets about it, doesn't mind me too much.
Huge companies who used to have a lot of good will putting out stuff like this, seemingly with absolutely zero reviews before hitting publish? What are they doing? Have everyone decided to just give up and give in to the slop? We need "engineering" to make a comeback.
I'm mostly concerned that something we used to see as a part of basic "software engineering" (verify that what you build is actually doing what you think it is) has suddenly made a very quick exit from the scene, in chase of outputting more LOC which is completely backwards.
This is also what I ask our engineers to do, but it's getting hard to enforce.
Vibing is incompatible with engineering and this practice is disgusting and NOT acceptable.
Perhaps usage of AI is a performance target he's being judged against, like at many tech companies today.
I'm starting to believe they are all right, actually. Maybe frontier models surpassed most humans, but the bar we should have for humans is really really low. I genuinely believe most people cannot distinguish llms capabilities from their own capabilities, and their are not wrong from the perspective they have.
How could you perceive, out in the wild, an essence that scapes you?
I have yet to see a counter-example
Even if Blockchain has tremendous impact, even if transformers are incredible (really) technology, even if NFTs could solve real world problems...you could basically say the same thing and be right, rounding up, 100% of the time, about anything technology related (and everything else as well). This truly is a clown world, but it is illegal to challenge it (or considered bad faith around here)
The project itself did compile most of the time it was being developed - the coding agents had been compiling it the whole time they were running on it.
Shortly after the blog post they updated the GitHub repo with compilation instructions and it worked. I took this screenshot with it: https://static.simonwillison.net/static/2026/cursor-simonwil...
The "it didn't even compile" criticism is valid in pointing out that they messed up the initial release, but if you think "it never compiled" you have an incorrect mental model.
If I install an Arch Linux, I don't say I 'installed Linux from scratch'.
I'd estimate that's a lot less than 60% of the "actual work" though.