Ask HN: Why all the sudden people are writing browsers with AI?

Because this is a very large and complex project it is specified very clearly, to the extent that the three-word prompt “build a browser” contains a large amount of detail.Similar to “building space invaders”, this is another useful test prompt to see how well an LLM can do a task of moderate complexity without a lot of instruction.

Recently I called building a browser a “Hello World” of complex parallel agent coding harnesses: https://simonwillison.net/2026/Jan/23/fastrender/#a-single-e…

I’m not sure any of these are good test signals for tasks of moderate complexity. Many solutions have already been included in the training data!You can create a web browser for Firefox trivially by copying and compiling the code, no transformer required.

But how do you define innovation, or indeed measure it?

It’s not that difficult to take an existing mature codebase and transform it in such a way that it looks significantly different but remains functionally unchanged.

This is a very different task than building something that hasn’t been built before.

I think a lot of people here are concerned that there are only three browsers now, and there are actually only two: Google Chrome and Safari for iOS. Internet Explorer is now just Chrome and it is still much smaller in usage statistics than when I use Firefox. I don’t consider things like Epiphany/Gnome Web to be serious contenders, and even if I did it would just be another niche browser. In my opinion, things like Netsurf don’t count without JS support.

It’s a little scary when a single megacorp has so much power over something as ostensibly open as the Internet, but it’s also taken historically incredible amounts of resources to build a browser, making it hard for new players to enter. A modern web browser is arguably more complex than an operating system… hell, it sure is Is An operating system. It touches many aspects of computer science, and maintaining web standards requires a lot of dedicated staff.

Because this has been a difficult problem for so long, it is an extremely attractive target when circumstances have changed. Here anyone can get a metaphorical intern to work as many hours as they want for basically ~$20-$100/month. A problem that would have been impossible for any one person five years ago suddenly seems “almost possible” when you can work at a higher level and have the annoying “code” details taken care of for you.

How does your custom AI-built browser challenge the current browser trinity?

LLM is trained on the code of existing browsers. It’s essentially a comprehensive process of converting code you don’t understand into code you don’t understand.

Code description is almost all description except protocols and standards.

If you understand the codebase of existing browsers (or at least can be confident in making arbitrary changes to existing browsers, perhaps with AI assistance?) then Tripoli won’t be dangerous since you can patch out manifestv3 whenever you want.

People also have the problem that they don’t test their websites to be compatible with your custom browser. But I would say this is a problem at the protocol level.

Blame Simon Willison 😉

“A common complaint from AI coding skeptics today is that LLMs are fine for toy projects, but cannot be used for anything larger and serious.

I think within 3 years it will be widely proven wrong, to the point where it won’t even be controversial anymore.

I chose a web browser here because much of the work of building a browser involves writing code that has to conform to a huge and challenging selection of both formal tests and unofficial websites.

Coding agents are really good at tasks where you can define a concrete goal and then set them to work toward that end.

A web browser is the most ambitious project I can think of that is based on those capabilities.”

https://simonwillison.net/2026/Jan/8/llm-predictions-for-202…

“The browser and this project were co-developed and very symbiotic, simply because the browser was a very useful objective for us to measure and iterate the progress of the harness. The goal was to iterate and research the multi-agent harness – the browser was just the research example or objective.”

https://simonwillison.net/2026/Jan/23/fastrender/

> Coding agents are really good at tasks where you can define a concrete goal and then set them to work towards that.

Especially those that are in the training data.

>A web browser is the most ambitious project I can think of that is based on those capabilities.”

I believe Linux and GCC are in the training data, so additional options may be OS and compiler.

> Coding agents are really good at tasks where you can define a concrete goal and then set them to work towards that.

Based entirely on other people’s work. Which is fine.

IMO it is a very verifiably correct and complex software. There are all kinds of test suites, but none is better than comparing a page rendered in Chrome.

Additionally, and perhaps more importantly, a browser is the perfect entry point for a lot of businesses.

For example, if you’re hoping to create an agent that people trust to manage their calendars, email, book their vacations, etc., browser is probably the perfect form factor for it.

Or if you’re an AI lab and want to use training data behind logged in websites – a browser that everyone uses is ultimately what you need.

There’s no need to pay millions of dollars to hire Chromium engineers from Google if you can just build and maintain a browser with agents.

The list goes on!

I would really like a pure RSS browser. Just good articles in standard format, no ads, no clickbait.

RSS readers are good but searching is still difficult – looking for a feed is not much fun.

However, I’ve just discovered fiddle, which looks half-way to what I want! https://feedle.world/

I want fiddle in a nice native app.

Because they want to sell AI and browsers are things users spend a lot of time on.

Although I don’t really see any added value there. I watch it for discovery and in-depth research but rarely watch it for regular browsing.

It would be very useful if AI could do this (although I doubt anything high-performance can be created this way). It is now economically unviable for humans to develop a browser engine from scratch. Microsoft and Opera gave up. Apple will probably do this someday too.

It’s the most complex piece of software on your computer (at this point it’s basically an entire sandbox operating system), the standards it has to adhere to are increasing every day, its performance optimization is significant yet adversarial (ie website owners have no incentive to make their sites efficient – the browser will be blamed for slowness, not them), and it doesn’t cost the user anything. No company can afford to maintain this unless it is offering a comprehensive strategy that will make billions.

This is a hard, big problem – but there are very comprehensive tests available and LLMs have a good understanding of it. You can also check out earlier automated ports of the JustHTML parser. A rendering engine is the next (admittedly bigger) step.

I say this as someone who also did this task about a week before seeing the cursor attempt.

Because they can. It’s fun.

AI makes it possible to do things that you would never have been able to do before, perhaps because of skill level or perhaps because of the time investment required.

It’s a lot of fun to create software that you could only dream about creating before AI.

Because someone tried it and it vaguely worked.

And browsers are one of those products where if it were easy to change Everyone will want different setups and features. Like the earlier Notes app.

Naturally they are all incomplete implementations because the AI ​​agent is mostly reusing open source components and for the stuff he decides to write himself, he doesn’t have the training data to implement a full rendering engine in one-shot.

This is why suddenly people are rolling their own sandboxes using AI (without even seeing the code!):

   Hype (for VCs), not invented here syndrome (NIH), peformative reasons for the impression of progress or just because they can.

On the other hand, motivating people to use it…

>I’ve seen a lot of frustration with validating the usefulness of AI. I guess to some extent from people who are trying to make money from it and to some extent from people who need a cognitive shortcut to get the work and creativity out of it.

I’ve seen a lot of frustration with the denial of AI’s utility. I guess to some extent from those who feel threatened by it and to some extent from others who are not keen on learning how to use it effectively.

(Sorry, couldn’t resist.)