Similar to "build space invaders", another useful test prompt for seeing how well an LLM can do at a medium complexity task without having to give it a great deal of instruction.
I called building a browser the "hello world" of complex parallel agent coding harnesses the other day: https://simonwillison.net/2026/Jan/23/fastrender/#a-single-e...
You can trivially produce a web browser by copying and compiling the code for firefox, no transformer needed.
The goal with these tests is to see if the models can make something new, not just copy an existing solution.
That is the goal, at least.
It's a little scary when a single megacorp has so much power over something ostensibly open like the internet, but it also has historically taken an incredible amount of resources in order to make a browser making it hard for new players to break in. A modern web browser is arguably more complicated than an operating system...hell, it arguably is an operating system. It touches a ton of aspects of computer science, and requires lots of dedicated workers to keep up with web standards.
Because it has been such an intractable problem for so long, it's an extremely tempting target when the circumstances have changed. Anyone here can basically have a metaphorical intern working as many hours as they want for ~$20-$100/month. A problem that would be impossible for a single person five years ago suddenly seems "almost possible" when you can work at a higher level and have the pesky "code" details taken care of for you.
It was already possible - The Ladybird browser started with one person and then a team of experts in maintaining browsers joined in. It makes a lot of sense for them to try to build one from scratch; with / without AI.
The problem which applies to all non mainstream browsers: Is anyone going to use your browser over the established ones because of some technical detail such as, who wrote it or what it is written in?
99.999% of people only care about whether if it just works and disappears away from the user whilst they're using it as a daily driver for browsing or doing work.
How is AI surely going to make this better the other than inexperienced folks throwing PRs they have not read? In fact, it helps the experts rather than the vibe coders.
Yes, by a person with prior browser-making experience and who was also building his own operating system. It's not that weird that people who know how to make browsers started making browsers.
Most people don't know how to make browsers, but doing so is still tempting. LLMs lower the barrier to entry tremendously.
To be clear, I'm not giving a value judgement to this; I'm not saying this is "good", just the why of it.
The LLMs are trained on the code of existing browsers. This is essentially a massive process of turning code you don't understand into code you don't understand.
The code details are pretty much all of the details, other than the protocols and standards.
If you understood the codebase of existing browsers (or at least could be confident in making arbitrary changes to existing browsers, perhaps with AI assistance?) then the triopoly wouldn't be threatening because you could just patch out manifestv3 whenever you want.
There is also the problem of people not testing their websites to be compatible with your custom browser. But I would say this is a problem on the protocol level.
As for training on existing browsers: they are trained on the whole corpus of the human thought process and can take many insights from other fields into the browser, they can write a browser in a completely new language without just transpiring but building it from first principles or as karparthy calls it spec based programming.
Prior to LLMs, creating a browser from scratch seemed like an insurmountable task for a single person. LLMs lower the barrier to entry, and it's a space that is tempting because it would be cool to be the one to create a new browser that people use.
“A common complaint today from AI coding skeptics is that LLMs are fine for toy projects but can’t be used for anything large and serious.
I think within 3 years that will be comprehensively proven incorrect, to the point that it won’t even be controversial anymore.
I picked a web browser here because so much of the work building a browser involves writing code that has to conform to an enormous and daunting selection of both formal tests and informal websites-in-the-wild.
Coding agents are really good at tasks where you can define a concrete goal and then set them to work iterating in that direction.
A web browser is the most ambitious project I can think of that leans into those capabilities.”
https://simonwillison.net/2026/Jan/8/llm-predictions-for-202...
“The browser and this project were co-developed and very symbiotic, only because the browser was a very useful objective for us to measure and iterate the progress of the harness. The goal was to iterate on and research the multi-agent harness—the browser was just the research example or objective.”
Wholly based on other people's work. Which is OK.
Specifically ones that are in the training data.
> A web browser is the most ambitious project I can think of that leans into those capabilities.”
I assume Linux and gcc are in the training data, so additional options may be OSes and compilers..
RSS readers are cool but the discovery is still tedious - looking for feeds isn't much fun.
I did just discover feedle though, which looks like halfway toward what I want! https://feedle.world/
I want feedle in a nice native app.
I wonder if Feedle will provide an API (even a paid API)? Could be nice to plug this into a metasearch engine as an additional data source. And I guess that might be necessary to add Feedle search to an RSS reader or native app as well.
https://www.zdnet.com/article/embrace-extend-extinguish-how-...
Also, and maybe more importantly, a browser is the kind of the perfect entry point for a lot of businesses.
If you’re for example hoping to build an agent that people rely on to manage their calendars, emails, book them vacations, etc.. a browser is probably the perfect form factor for that.
Or if you’re a AI lab and want to tap into training data behind logged in websites - a browser everyone uses is the ultimately what you need.
If you can build and maintain a browser from scratch with just agents then there’s no need to pay millions of dollars to hire away chromium engineers from Google.
The list goes on!
And browsers are one of those products where if it was simple to alter everyone would like different setups and features. Like the Notes app of before.
Naturally they are all incomplete implementations because the AI agent is mostly reusing open source components and for the stuff it decides to write itself, doesn't have the training data to implement a full rendering engine in a one-shot.
It's the most complex piece of software on your computer (it's basically an entire sandboxed operating system, at this point), the standards it has to adhere to are expanding every day, its performance optimization is critical yet adversarial (i.e. website owners have no incentive to make their sites efficient - the browser will be blamed for slowness, not them), and it costs the user nothing. No company can afford to maintain that, unless it's serving some broader strategy that's earning billions.
Is there? Who else? I thought I was the only one, besides Cursor's half-hearted attempt :P Where are the rest of the AI-made browsers? Now I want to see them.
In my case, I saw Cursor's blog post bragging about using hundreds of agents producing millions of lines of code, and the results were underwhelming (not to mention broken until a human stepped in). Initially I tried to dig into their codebase to understand how it got to where it was, but gave up.
Instead, I changed my approach, and decided to build a browser using only OS libraries in order to show that millions of LOC for what they got, is very underwhelming.
:)
AI makes it possible to do things you'd never have been able to do before, perhaps due to skill level or perhaps due to the time investment required.
It is a lot of fun to make software that you'd only been able to dream about making before AI.
I say this as someone who also took on this task, about a week before I saw Cursor's attempt.
What was the results? As far as I can tell, less agents work better, paired together with someone who knows solid engineering.
I burned through a week of codex max subscription quota with it making slow progress. Slower than I wanted, but consistent slow progress. Not long after though it hit a brick wall and the second week+ of quota did not really make much progress. It can churn for hours and get one test working, or churn for hours and fail.
I might give it a go with more guidance, because I actually really would like an html/css renderer with some specific properties about how I interact with it from code.
Hype (for VCs), not invented here syndrome (NIH), performative reasons for the impression of progress or just because they can.
I don't really see the added value there though. I do see it for search and deep research but much less for regular browsing.
(I didn't know that, actually)
I know the analogy is not perfect but it's the kind of project that wasn't feasible for a single dev before LLMs so now it just seems like a fun thing to try for some people.
- https://github.com/hiwavebrowser/hiwave
- https://github.com/wilsonzlin/fastrender
- https://github.com/embedding-shapes/one-agent-one-browser
I've started tracking them on this tag on my blog: https://simonwillison.net/tags/browser-challenge/
Ehh, really pushing my buttons to continue now huh? I have other things to do! twitching
Guess I'll start sketching out some lame JS engine from scratch too...
I've noticed a lot of desperation to deny the usefulness of AI. I guess partly from those who feel threatened by it and partly from others who aren't curious enough to figure out how to use it effectively.
(Sorry, couldn't resist.)