Show HN: One Human + One Agent = One Browser From Scratch in 20K LOC

https://emsh.cat/one-human-one-agent-one-browser/

324 146 6 months ago

I set some rules for myself: three days of total time, no 3rd party Rust crates, allowed to use commonly available OS libraries, has to support X11/Windows/macOS and can render some websites.

After three days, I have it working with around 20K LOC, whereas ~14K is the browser engine itself + X11, then 6K is just Windows+macOS support.

Source code + CI built binaries are available here if you wanna try it out: https://github.com/embedding-shapes/one-agent-one-browser

jacquesm4• 6 months ago

Those are excellent constraints.

chatmasta• 6 months ago

Did you use Claude code? How many tokens did you burn? What’d it cost? What model did you use?

embedding-shape42• 6 months ago

Codex, no idea about tokens, I'll upload the session data probably tomorrow so you could see exactly what was done. I pay ~200 EUR/month for the ChatGPT Pro plan, prorating days I guess it'll be ~19 EUR for three days. Model used for everything was gpt-5.2 with reasoning effort set to xhigh.

bhadass2• 6 months ago

very impressive!

it's amazing how far we've come in 20 years. i was a (very minor) contributor to khtml/konqueror (before apple got involved w/ webkit) in the early 2000s, and back then it was such a labor intensive process to even create a halfway working engine. like, months of work just to get basic rendering somewhat correct on a very small portion of the web (which was obv much smaller)

in addition to agentic coding, i think for this specific task having css-spec/html-spec/web-platform-tests as machine readable test suites helps a LOT. the agent can actually validate against real specs.

back in the day, despite having gecko as an open source reference, in practice the "standards" were whatever IE was doing. so you'd spend weeks implementing something only to discover every site was coded for IE's quirks lmao. for all of their other faults, google/apple and other contributors helped bring in discipline to that.

embedding-shape42• 6 months ago

i think for this specific task having css-spec/html-spec/web-platform-tests as machine readable test suites helps a LOT

You know, I placed the specs in the repository with that goal (even sneaked in a repo that needs compiling before being usable), but as far as I can see, the agent never actually peeked into that directory nor read anything from them in the end.

It'll be easier to see once I made all the agent sessions public, and I might be wrong (I didn't observe the agent at all times), but seems the agent never used though.

simonw14• 6 months ago

This is a notably better demonstration of a coding agent generated browser than Cursor's FastRender - it's a fraction of the size (20,000 lines of Rust compared to ~1.6m), uses way fewer dependencies (just system libraries for rendering images and text) and the code is actually quite readable - here's the flexbox implementation, for example: https://github.com/embedding-shapes/one-agent-one-browser/bl...

Here's my own screenshot of it rendering my blog - https://bsky.app/profile/simonwillison.net/post/3mdg2oo6bms2... - it handles the layout and CSS gradiants really well, renders the SVG feed icon but fails to render a PNG image.

I thought "build a browser that renders HTML+CSS" was the perfect task for demonstrating a massively parallel agent setup because it couldn't be productively achieved in a few thousand lines of code by a single coding agent. Turns out I was wrong!

vidarh3• 6 months ago

I think the human + agent thing absolutely will make a huge difference. I see regularly that Claude can totally off piste and eventually claw itself back with a proper agent setup but it will take a lot of time if I don't spot it and get it back on track.

I have one project Claude is working on right now where I'm testing a setup to attempt to take myself more out of the loop, because that is the hard part. It's "easy" to get an agent to multiply your output. It's hard to make that scale with your willingness to spend on tokens rather than with your ability to read and review and direct.

I've ended up with roughly this (it's nothing particularly special):

- Runs a evaluator that evaluates the current state and assigns scores across multiple metrics.

- If a given score is above a given threshold, expand the test suite automatically.

- If the score is below a given threshold, spawn a "research agent" that investgates why the scores don't meet expectations.

- The research agent delivers a report, that is passed to an implementation agent.

- The main agent re-runs the scoring, and if it doesn't show an improvement on one or more of the metrics, the commit is discarded, and notes made of what was tried, and why it failed.

It takes a bit of trial and error to get it right (e.g. "it's the test suite that is wrong" came up early, and the main agent was almost talked into revising the test suite to remove the "problematic" tests) but a division sort of like this lets Claude do more sensible stuff for me. Throwing away commits feels drastic - an option is to let it run a little cycle of commit -> evaluate -> redo a few times before the final judgement, maybe - but it so far it feels like it'll scale better. Less crap makes it into the project.

And I think this will work better than to treat these agents as if they are developers whose output costs 100x as much.

Code so cheap it is disposable should change the workflows.

So while I agree this is a better demonstration of a good way to build a browser, it's a less interesting demonstration as well. Now that we've seen people show that something like FastRender is possible, expect people to experiment with similarly ambitious projects but with more thought put into scoring/evaluation, including on code size and dependencies.

embedding-shape42• 6 months ago

I think the human + agent thing absolutely will make a huge difference.

Just the day(s) before, I was thinking about this too, and I think what will make the biggest difference is humans who posses "Good Taste". I wrote a bunch about it here: https://emsh.cat/good-taste/

I think the ending is most apt, and where I think we're going wrong right now:

I feel like we're building the wrong things. The whole vibe right now is "replace the human part" instead of "make better tools for the human part". I don't want a machine that replaces my taste, I want tools that help me use my taste better; see the cut faster, compare directions, compare architectural choices, find where I've missed things, catch when we're going into generics, and help me make sharper intentional choices.

queenkjuul2• 6 months ago

I'm confused, what did FastRender show is possible? That's cursor's agent-built browser right?

The one that people couldn't compile, and was largely a failed attempt to stitch together existing libraries?

Imustaskforhelp6• 6 months ago

To me I really like how embedding shapes took things in his own hands and actually built it. It really proved a point at such a scale where I don't think any recent example can point to.

It's great to see hackernews be so core part of it haha.

I thought "build a browser that renders HTML+CSS" was the perfect task for demonstrating a massively parallel agent setup because it couldn't be productively achieved in a few thousand lines of code by a single coding agent. Turns out I was wrong!

I do wonder if tech people from future/present are gonna witness this as a goliath vs david story. 20k 1 human 1 agent beats 5 million$ 1.6 millions loc browser changing how even the massive AI users/pioneers at the time thought about the use of AI

Looks like I have watched some documentaries recently but why do I feel like a documentary about this whole thing can be created in future.

But also, More and more I am feeling like AI is an absolute black box, nobody knows how to do things but we are all kind of doing experiments with it and seeing what sticks (like how we now have definitive proof that 1 human 1 agent > many agents no human in the loop)

And this is when we are 1 month in 2026, who knows what other experiments and proofs happen this year to find more about this black box, and about its usefulness or not.

Simon, it would be interesting if you could read the thread of predictions of 2026 thread in hn each month or quaterly to see how many people were wrong or right about AI as we figure out more things perhaps.

g947o• 6 months ago

I think most people would agree that this is much more superior than Cursor's "browser" from an engineering perspective -- it doesn't do much but does it well, as you pointed out.

What it tells me is that "effectively using agents" can be much more important than just throwing tokens at a problem and see what comes out. I myself have completely deleted several small vibe-coded projects without even going over the code, because what often happens is that, two days after the code is generated, I realize that I was solving the wrong problem or using the wrong approach.

A coding agent doesn't care. It most likely just does whatever you ask it to do with no pushback. While in some cases it's worth using them to validate an idea, often you dig a deeper hole for yourself if you go down a wrong path in the first place.

embedding-shape42• 6 months ago

Yeah, I agree with all of what you wrote, how these are used seems (to me) to be more important than how they're built. If you don't know software engineering, a software engineering agent isn't suddenly gonna make you one, but someone who already knows the craft, can be very effective with one.

Amplifiers, rather than replacements. I think the community at large still thinks LLMs and agents are gonna be "replacing" knowledge, which I think is far from the truth.

rahimnathwani• 6 months ago

This is awesome. Would you be willing to share more about your prompts? I'm particularly interested in how you prompted it to get the first few things working.

embedding-shape42• 6 months ago

Yes, I'm currently putting it all together and will make it public via the blog post. Just need to go through all of it first to ensure nothing secret/private leaks, will update once I've made it public.

jacquesm4• 6 months ago

This post is far more interesting than many others on the same subject, not because of what is built but because of how it it is built. There is a ton of noise on this subject and most of it seems to focus on the thing - or even on the author - rather than on the process, the constraints and the outcome.

embedding-shape42• 6 months ago

Thanks, means a lot. As the author of one such article (that might have been the catalyst even), I'm guilty of this myself, and as I dove deeper into understanding what Cursor actually built, and what they think was the "success", the less sense everything made to me.

That's why taking a step back and seeing what's actually hard in the process and bad with the output, felt like it made more sense to chase after, rather than anything else.

jacquesm4• 6 months ago

I think the Cursor example is as bad as it gets and this is as good as it gets.

FWIW I ran your binary and was pleasantly surprised, but my low expectations probably helped ;)

rvz• 6 months ago

I'm going to upgrade my prediction for 2029: I think we're going to get a production-grade web browser built by a small team using AI assistance by then.

That is Ladybird Browser if that was not already obvious.

dewey• 6 months ago

For the curious, they have a reasonable AI policy:

https://github.com/LadybirdBrowser/ladybird/blob/master/CONT...

simonw14• 6 months ago

Ladybird (a project I deeply respect) had a several year head start.

mwcampbell• 6 months ago

Impressive work.

I wonder if you've looked into what it would take to implement accessibility while maintaining your no-Rust-dependencies rule. On Windows and macOS, it's straightforward enough to implement UI Automation and the Cocoa NSAccessibility protocols respectively. On Unix/X11, as I see it, your options are:

1. Implement AT-SPI with a new from-scratch D-Bus implementation.

2. Implement AT-SPI with one of the D-Bus C libraries (GLib, libdbus, or sdbus).

3. Use GTK, or maybe Qt.

QuadmasterXLII2• 6 months ago

The rendering is pretty chaotic when I tried it- not that far off from just the text in the html tags, in some size, color, and placement on the screen. This sounds like unfairness, but there is some motte-and-bailey where if you claim to be a browser, I get to evaluate on stuff like links being consistently blue and underlined ( as is, they are sometimes blue and sometimes underlined, without a clear pattern- if they were never formatted differently from standard text, I would just buy this as a feature not implemented yet). It may be that some of the rendering is not supported on windows- the back button certainly isn't. I guess if I want to make my criticism actually legitimate I should make a "one human and no agent browser" post that just regexes out stuff that looks like content and formats it at random. The binary I downloaded definitely overperforms at the hacker news homepage and simonw's blog.

embedding-shape42• 6 months ago

It's a really basic browser. It's made less as an independent thing, and more as a reply to https://cursor.com/blog/scaling-agents, so as long as it does more or less the same as theirs, but is less LOC, it does what I set out for it to do :)

I get to evaluate on stuff like links being consistently blue and underlined

Yeah, this browser doesn't have a "default stylesheet" like a regular browser. Probably should have added that, but was mostly just curious about rendering the websites from the web, rather than using what browsers think the web should look like.

It may be that some of the rendering is not supported on windows- the back button certainly isn't.

Hmm, on Windows 11 the back button should definitively work, tried that just last night. Are you perhaps on Windows 10? I have not tried that myself, should work but might be why.

QuadmasterXLII2• 6 months ago

It is both extraordinarily impressive in an absolute sense, and fairly disappointing specifically comparing my result on a a random smattering of other no-js websites, to the expectation I had from the simonw screenshot (which to be clear is not an expectation you had control over, as you are not simonw). I'm familiar with this pattern from all the rest of my trying frontier ML results!

Yep, I ran it on an old windows 10 VM I had puttering about.

I think it must have a default link styling somewhere, as some links are the classic blue that as far as I know I intentionally styled to be black- but this could be css spaghetti in tufte.css finally coming to haunt me.

storystarling6• 6 months ago

How did you handle the context window for 20k lines? I assume you aren't feeding the whole codebase in every time given the API costs. I've struggled to keep agents coherent on larger projects without blowing the budget, so I'm curious if you used a specific scoping strategy here.

simonw14• 6 months ago

GPT-5.2 has a 400,000 token context window. Claude Opus 4.5 is just 200,000 tokens. To my surprise this doesn't seem to limit their ability to work with much larger codebases - the coding agent harnesses have got really good at grepping for just the code that they need to have in-context, similar to how a human engineer can make changes to a million lines of code without having to hold it all in their head at once.

storystarling6• 6 months ago

That explains the coherence, but I'm curious about the mechanics of the retrieval. Is it AST-based to map dependencies or are you just using vector search? I assume you still have to filter pretty aggressively to keep the token costs viable for a commercial tool.

nurettin2• 6 months ago

You don't load the entire project into the context. You let the agent work on a few 600-800 line files one feature at a time.

storystarling6• 6 months ago

Right, but how does it know which files to pick? I'm curious if you're using a dependency graph or embeddings for that discovery step, since getting the agent to self-select the right scope is usually the main bottleneck.

embedding-shape42• 6 months ago

I didn't, Codex (tui/cli) did, it does it all by itself. I have one REQUIREMENTS.md which is specific to the project, a AGENTS.md that I reuse across most projects, then I give Codex (gpt-5.2 with reasoning effort set to xhigh) a prompt + screenshot, tells it to get it to work somewhat similar, waits until it completes, reviewed that it worked, then continued.

Most of the time when I develop professionally, I restart the session after each successful change, for this project, I initially tried to let one session go as long as possible, but eventually I reverted back to my old behavior of restarting from 0 after successful changes.

For knowing what file it should read/write, it uses `ls`, `tree` and `ag ` most commonly, there is no out-of-band indexing or anything, just a unix shell controlled by a LLM via tool calls.

pulkas• 6 months ago

The Mythical Man-Month, revisited

lelele• 6 months ago

What do you mean?

embedding-shape42• 6 months ago

I think it means they'd like to have a baby with me, and the more agents we can add, the faster the baby can incubate. Usual stuff :)

Imustaskforhelp6• 6 months ago

I feel like I have talked to Embedding-shape on Hackernews quite a lot that I recognize him. So it was a proud like moment when I saw his hackernews & github comments on a youtube video [0]about the recent cursor thing

It's great to see him make this. I didn't know that he had a blog but looks good to me. Bookmarked now.

I feel like although Cursor burned 5 million$, we saw that and now Embedding shapes takeaway

If one person with one agent can produce equal or better results than "hundreds of agents for weeks", then the answer to the question: "Can we scale autonomous coding by throwing more agents at a problem?", probably has a more pessimistic answer than some expected.

Effectively to me this feels like answering the query which was being what if we have thousands of AI agents who can build a complex project autonomously with no Human. That idea seems dead now. Humans being in the loop will have a much higher productivity and end result.

I feel like the lure behind the Cursor project was to find if its able to replace humans completely in a extremely large project and the answer's right now no (and I have a feeling [bias?] that the answer's gonna stay that way)

Emsh I have a question tho, can you tell me about your background if possible? Have you been involved in browser development or any related endeavours or was this a first new one for you? From what I can feel/have talked with you, I do feel like the answer's yes that you have worked in browser space but I am still curious to know the answer.

A question which is coming to my mind is how much would be the difference between 1 expert human 1 agent and 1 (non expert) say Junior dev human 1 agent and 1 completely non expert say a normal person/less techie person 1 agent go?

What are your guys prediction on it?

How would the economics of becoming an "expert" or becoming a jack of all trades (junior dev) in a field fare with this new technology/toy that we got.

how much productivity gains could be from 1 non expert -> junior dev and the same question for junior -> senior dev in this particular context

[0] Cursor Is Lying To Developers… : https://www.youtube.com/watch?v=U7s_CaI93Mo

simonw14• 6 months ago

I don't think the Cursor thing was about replacing humans entirely.

(If it was that's bad news for them as a company that sells tools to human developers!)

It was about scaling coding agents up to much larger projects by coordinating and running them in parallel. They chose a web browser for that not because they wanted to build a web browser, but because it seemed like the ideal example of a well specified but enormous (million line+) project which multiple parallel agents could take on where a single agent wouldn't be able to make progress.

embedding-shape's project here disproves that last bit - that you need parallel agents to build a competent web renderer - by achieving a more impressive result with just one Codex agent in a few days.

Imustaskforhelp6• 6 months ago

I don't think the Cursor thing was about replacing humans entirely.

I think how I saw things was that somehow Cursor was/is still targetted very heavily on vibe coding in a similar fashion of bolt.dev or lovable and I even saw some vibe coders youtube try to see the difference and honestly at the end Cursor had a preferable pricing than the other two and that's how I felt Cursor was.

Of course Cursor's for the more techie person as well but I feel as if they would shift more and more towards Claude Code or similar which are subsidized by the provider (Anthropic) itself, something not possible for Cursor to do unless burning big B's which it already has done.

So Cursor's growth was definitely towards the more vibe coders side.

Now coming to my main point which is that I had the feeling that what cursor was trying to achieve wasn't trying to replace humans entirely but replace humans from the loop Aka Vibe coding. Instead of having engineers, if suppose the Cursor experiment was sucessful, the idea (which people felt when it was first released instantly) was that the engineering itself would've been dead & instead the jobs would've turned into management from a bird's eye view (not managing agent's individually or being aware of what they did or being in any capacity within the loop)

I feel like this might've been the reason they were willing to burn 5 million$ for.

If you could've been able to convince engineers considering browsers are taken as the holy grail of hardness that they are better off being managers, then a vibe coding product like Cursor would be really lucrative.

Atleast that's my understanding, I can be wrong I usually am and I don't have anything against Cursor. (I actually used to use Cursor earlier)

But the embedding shapes project shows that engineering is very much still alive and beneficial net. He produced a better result with very minimal costs than 5 million$ inference costs project.

embedding-shape's project here disproves that last bit - that you need parallel agents to build a competent web renderer - by achieving a more impressive result with just one Codex agent in a few days.

Simon, I think that browsers got the idea of this autonomous agents partially because of your really famous post about how independent tests can lead to easier ports via agents. Browsers have a lot of independent tests.

So Simon, perhaps I may have over-generalized but do you know of any ideas where the idea of parallel agents is actually good now that browsers are off, personally after this project, I can't really think of any. When the Cursor thing first launched or when I first heard of it recently, I thought that browsers did make sense for some reason but now that its out of the window, I am not sure if there are any other projects where massively parallel agents might be even net positive over 1 human + 1 agent as Emsh.

forgotpwd163• 6 months ago

Impressive. Very nice. (Let's see Paul Allen's browser. /s) Can say is Brooks's law in action. What one human and one agent can do in 3d, one human and hundreds of agents can do in few weeks. A modern retake of the old joke.

without using any 3rd party libraries

Seems to be an easier for coding agents to implement from scratch over using libraries.

avmich• 6 months ago

Next thing would probably be an OS. With different APIs, the browser could be not constrained by existing standards. Generation of a good set of applications making working in the OS convenient - starting with GNU set? And then we can approach CPU architecture - again, without constraint to existing languages or instruction sets. That should be interesting to play with.

forgotpwd163• 6 months ago

Someone has already done this: https://github.com/viralcode/vib-OS

Also, someone made a similar comment not too long ago. So people surely are curious if this is possible. Kinda surprised this project's submission didn't got popular.

happytoexplain2• 6 months ago

What kind of time frame do you ballpark this would have taken you on your own?

I know it's a little apples-and-oranges (you and the agent wouldn't produce the exact same thing), but I'm not asking because I'm interested in the man-hour savings. Rather, I want to get a perspective on what kind of expertise went into the guidance (without having to read all the guidance and be familiar with browser implementation myself). "How long this would have taken the author" seems like one possible proxy for "how much pre-existing experience went into this agent's guidance".

simonw14• 6 months ago

I have a fun little tool which runs the year-2000-era sloccount algorithm (which is Perl and C so I run it in WebAssembly) to estimate the time and cost of a project here: https://tools.simonwillison.net/sloccount

If you paste https://github.com/embedding-shapes/one-agent-one-browser into the "GitHub Repository" tab it estimates 4.58 person-years and $618,599 by year-2000 standards, or 5.61 years and $1,381,079 according to my very non-trustworthy 2025 estimate upgrade.

pizlonator3• 6 months ago

I pasted a subset of the Fil-C source code into your tool and it says 6 person years. I just pasted the compiler pass and the obvious parts of the runtime.

Note that I started the project in Nov 2023 and can only work on it maybe 1-2 hours a day because it's just a side project.

So I think your tool either estimates based on very bad programmers, or it's just wrong. Or maybe 10x programmers are real and I am him

embedding-shape42• 6 months ago

What kind of time frame do you ballpark this would have taken you on your own?

I don't think I'd be able to do this on my own. Not that I don't know Rust, but because I don't know X11 (nor macOS or Windows) well enough to even know where to begin.

I've been a Linux user for almost two decades, so I know my way around my system, but never developed X11 applications or anything, I'm mostly a web developer who jumped around various roles through the years. Spent a lot of time caring deeply about testing, infrastructure, architecture/design and communication between humans, might have given me a slight edge in programming together with agents.

happytoexplain2• 6 months ago

Hmm, well I'm more interested in the browser part rather than the windowing part - I feel like it makes more sense that LLMs can be somewhat competent with windowing frameworks even if the prompter is not super experienced. Regardless, there's probably not a concise way to get what I'm looking for - instead, I'm looking forward to seeing your config/input! I'm super curious.

barredo• 6 months ago

The binaries are only around 1 MB for Linux, Mac and Windows. Very impressive https://github.com/embedding-shapes/one-agent-one-browser/re...

embedding-shape42• 6 months ago

Fun fact, not until someone mentioned how small the binaries did I notice! Fun little side-effect from the various constraints and requirements I set in the REQUIREMENTS.md I suppose.

userbinator• 6 months ago

"only around 1MB" is not particularly impressive in absolute terms... there are a few browsers which are the same or smaller, and yet more functional.

https://tinyapps.org/network.html

Of course, "AI-generated browser is 1MB" is neither here nor there.

embedding-shape42• 6 months ago

Tried building with some other arguments/configs, went from 1,2M on X11 to 664K, which seems to place it under Lynx (text-only, 714k) but above OffByOne (full HTML 3.2, 409k). Of course, my experiment barely implements anything from a "real" browser, so unfair comparison really.

Neat collection of apps nonetheless, some really impressive stuff in there.

simonw14• 6 months ago

Do any of those handle CSS and SVG?

hedgehog4• 6 months ago

This looks pretty solid. I think you can make this process more efficient by decomposing the problem into layers that are more easily testable, e.g. testing topological relationships of DOM elements after parse, then spatial after layout, then eventually pixels on things like ACID2 or whatever the modern equivalent is. The models can often come up with tests more accurately than they get the code right the first time. There are often also invariants that can be used to identify bugs without ground truth, e.g rendering the page with slightly different widths you can make some assertions about how far elements will move.

embedding-shape42• 6 months ago

There are often also invariants that can be used to identify bugs without ground truth, e.g rendering the page with slightly different widths you can make some assertions about how far elements will move.

That's really interesting and sounds useful! I'm wondering if there are general guidelines/requirements (not specific to browsers) that could kind of "trigger" those things in the agent, without explicitly telling it. I think generally that's how I try to approach prompting.

hedgehog4• 6 months ago

I think if you explain that general idea the models can figure it enough to write into an implementation plan, at least some of the time. Interesting problem though.

socalgal22• 6 months ago

the modern equivalent is the Web Platform Tests

https://web-platform-tests.org/

hedgehog4• 6 months ago

Amazing. I think if I were taking on the build-a-browser project I would pair that with the WhatWG HTML spec to come up with a task list (based on the spec line-by-line) linked to specific tests associated with each task. Then of course need an overall architecture and behavioral spec for how the browser part behaves beyond just rendering. A developer steering process full time might be able to get within 80% parity of existing browsers in a month. It would be an interesting experiment.

sosodev• 6 months ago

The browser works shockingly well considering it was created in 72 hours. It can render Wikipedia well enough to read and browse articles. With some basic form handling and browser standards (url bar, history, bookmarks, etc) it would be a viable way to consume text based content.

embedding-shape42• 6 months ago

I can't say my fingers (codex's fingers) haven't been itching to add some small features which would basically make it a viable browser for myself at least, for 90% of my browsing.

But I think this is one of those experiments that I need to put a halt to sooner rather than later, because the scope can always grow, my mind really likes those sorts of projects, and I don't have the time for that right now :)

GaggiX2• 6 months ago

It would be really cool if it was able to render Wikipedia correctly, I really like the idea of a browser with minimal dependencies having the ability to navigate most static websites, this one for now compiles instantly and it's incredibly small.

fabrice_d• 6 months ago

This is a cool project, and to render Simon's blog will likely become the #1 goal of AI produced "web browsers".

But we're very far from a browser here, so that's not that impressive. Writing a basic renderer is really not that hard, and matches the effort and low LoC from that experiment. This is similar to countless graphical toolkits that have been written since the 70s.

I know Servo has a "no AI contribution" policy, but I still would be more impressed by a Servo fork that gets missing APIs implemented by an AI, with WPT tests passing etc. It's a lot less marketable I guess. Go add something like WebTransport for instance, it's a recent API so the spec should be properly written and there's a good test suite.

Dave3of5• 6 months ago

100% agree this isn't a browser. It's better than the previous attempt but fails to render even basic html websites correctly and crashes constantly.

The fact that it compiles is better the the cursor dude. "It Compiles" is a very low bar to working software.

embedding-shape42• 6 months ago

I think what I wanted to demonstrate here was less "You can build a browser with an agent", and more how bullshit Cursor's initial claim was, that "hundreds of agents" somehow managed to build something good, autonomously. It's more of a continuation of a blog post I wrote some days ago (https://emsh.cat/cursor-implied-success-without-evidence/), than a standalone proof of "agents can build browsers".

Unfortunately, this context is kind of implicit, I don't actually mention it in the blog post, which I probably should have done, that's my fault.

deadbabe5• 6 months ago

This is not that impressive, there are numerous examples of browsers for training data to reference.

embedding-shape42• 6 months ago

Damn, ok, what should I attempt instead, that could impress even you?

anonymous908213• 6 months ago

Actually good software that is suitable for mass adoption would go a long way to convincing a lot of people. This is just, yet another, proof-of-concept. Something which LLMs obviously can do, and which never seems to translate to real-world software people use. Parsing and rendering text is really not the hard part of building a browser, and there's no telling how closely the code mirrors existing open-source implementations if you aren't versed on the subject.

That said, I think some credit is due. This is still a nice weekend project as far as LLMs go, and I respect that you had a specific goal in mind (showing a better approach than Cursor's nonsense, that gets better results in less time with less cost) and achieved it quickly and decisively. It has not really changed my priors on LLMs in any way, though. If anything it just confirms them, particularly that the "agent swarm" stuff is a complete non-starter and demonstrates how ridiculous that avenue of hype is.

simonw14• 6 months ago

I don't buy this.

It implies that the agents could only do this because they could regurgitate previous browsers from their training data.

Anyone who's watched a coding agent work will see why that's unlikely to be what's happening. If that's all they were doing, why did it take three days and thousands of changes and tool calls to get to a working result?

I also know that AI labs treat regurgitation of training data as a bug and invest a lot of effort into making it unlikely to happen.

I recommend avoiding the temptation to look at things like this and say "yeah, that's not impressive, it saw that in the training data already". It's not a useful mental model to hold.

deadbabe5• 6 months ago

It took three days because... agents suck.

But yes, with enough prodding they will eventually build you something that's been built before. Don't see why that's particularly impressive. It's in the training data.

usef-• 6 months ago

What would be impressive to you?

deadbabe5• 6 months ago

A browser so unique and strange it is literally unlike anything we've ever seen to date, using entirely new UI patterns and paradigms.

tonyhart7• 6 months ago

one human
one agent
one browser
one million nvidia gpu

embedding-shape42• 6 months ago

Next time I'll do it on my GPU, then it'll be using just a 10K GPU, that's fine right?

madmaniak• 6 months ago

But when I install Firefox or Chrome it's much faster, much better and also someone else's code. Also copied and pasted by machine. Just I don't claim it's mine.

aix12• 6 months ago

Functionality aside, I'd find it very interesting to see a security audit of a code base like this.

I searched for "security" and "vuln" in both the article and this discussion thread, and found no matches.

I guess the code being in Rust helps, but to what exent can one just rely on guarantees provided by the language?

(I know practically nothing about Rust.)

embedding-shape42• 6 months ago

Hah, yeah, zero regards to security, don't run this without sandbox and load arbitrary websites :)

I don't think Rust helps much except preventing some very basic issues, for example, I don't think it even checks that URLs aren't referencing local files on disk, who knows how the path handling works, might be able to put absolute paths on remote pages and load local content? Unsure, but wouldn't surprise me.

Might be a bit safer due to no JS engine, so even if someone did what I outlined before, they couldn't really exfiltrate anything, there is no POST/PUT requests or forms or anything :)

I'm sure if someone did a proper audit they'd find double-digit high severity issues, at least.

dvrp• 6 months ago

It's interesting to think that—independently of what you think of Cursor's browser implementation being truly "from scratch" or not—the fact that people are implementing browsers from scratch with agents happened because of Cursor's post. In other words, in a twisted and funny way, this browser exists because of Cursor's agent.

This is how we should be thinking about AI safety!

embedding-shape42• 6 months ago

I mean I wanted to demonstrate further how wrong and misleading I think their initial blog post was so yeah, I made this because of what they said and marketed :)

socalgal22• 6 months ago

I'm having a hard time imagining how 20k lines of code gets a browser with no libraries. Just zlib by itself is 12k lines. freetype is 30k lines, or stb_truetype is 5k lines. Something doesn't seem like it's adding up. Am I missing something? Is this just calling into the OS for rendering?

embedding-shape42• 6 months ago

No Rust dependencies. Commonly available system libraries/frameworks are used for the actual drawing. The README of the repository outlines exactly which ones are being used on what OS.

senko• 6 months ago

On Linux, there are 78 dynamically linked libraries, such as for X11, vector graphics, glib/gobjectlibgobject, graphics formats, crypto, encryption, etc.

mdavid6263• 6 months ago

What’s the point of this?

embedding-shape42• 6 months ago

What's the point of anything really?

A more real answer: Read the first 6 words of the submission article.

mdavid6263• 6 months ago

It feels like this mentality is taking over the world. Porn instead of sex, short videos instead of real life interactions, AI generated code instead of software engineering, sugar/chemicals instead of food and so on...

All, just to have fun.

Very sad.

nenadg• 6 months ago

(no JS tho)

this is a feature

jFriedensreich• 6 months ago

My community and me are waiting for browserBench for a while now and happy to see it finally starting. Browsers are arguably one of the most complex and foundational piece of software, the ability to create something like this from scratch will be an important evaluation as limits of what is possible are harder and harder to find.

micimize• 6 months ago

An obvious nice thing here compared to the cursor post is the human involvement gives some minimum threshold confidence that the writer of the post has actually verified the claims they've made :^) Illustrates how human comprehension is itself a valuable "artifact" we won't soon be able to write off.

My comment on the cursor post for context: https://news.ycombinator.com/item?id=46625491

polyglotfacto3• 6 months ago

This one's really nice.

- clear code structure and good architecture(modular approach reminiscent of Blitz but not as radical, like Blitz-lite).

- Very easy to follow the code and understand how the main render loop works:

    - For Mac: main loop is at https://github.com/embedding-shapes/one-agent-one-browser/blob/master/src/platform/macos/windowed.rs#L74
   
    - You can see clearly how UI events as passed to the App to handle. 

    - App::tick allows the app to handle internal events(Servoshell does something similar with `spin_event_loop` at https://github.com/servo/servo/blob/611f3ef1625f4972337c247521f3a1d65040bd56/components/servo/servo.rs#L176)

    - If a redraw is needed, the main render logic is at https://github.com/embedding-shapes/one-agent-one-browser/blob/master/src/platform/macos/windowed.rs#L221 and calls into `render` of App, which computes a display list(layout) and then translates it into commands to the generic painter, which internally turns those into platform specific graphics operations.

- It's interesting how the painter for Mac uses Cocoa for graphics; very different from Servo which uses Webrender or Blitz which(in some path) uses Vello(itself using wgpu). I'd say using Cocoa like that might be closer to what React-Native does(expert to comfirm this pls?). Btw this kind of platform specific bindings is a strength of AI coding(and a real pain to do by hand).

- Nice modularity between the platform and browser app parts achieved with the App and Painter traits.

How to improve it further? I'd say try to map how the architecture correspond to Web standards, such as https://html.spec.whatwg.org/multipage/webappapis.html#event...

Wouldn't have to be precise and comprehensive, but for example parts of App::tick could be documented as an initial attempt to implement a part of the web event-loop and `render` as an attempt at implementing the update-the-rendering task.

You could also split the web engine part from the app embedding it in a similar way to the current split between platform and app.

Far superior, and more cost effective, than the attempt at scaling autonomous agent coding pursued by Fastrender. Shows how the important part isn't how many agents you can run in parallel, but rather how good of an idea the human overseeing the project has(or rather: develops).

embedding-shape42• 6 months ago

Hey, thanks a bunch of the review of the code itself! Personally I hadn't really looked deeply at it yet myself, especially the Windows and macOS code, interesting to hear that it as a slightly different approach for the painter, to me that's slightly unexpected.

Agree with your conclusion :)

Would be interesting to see how the architecture/design would change if I had focused the agent on making as modular and reusable code as possible, as currently I had some constraints about it, but not too strict. You bring up lots of interesting points, thanks again!

polyglotfacto3• 6 months ago

you're welcome.

Show HN: One Human + One Agent = One Browser From Scratch in 20K LOC

Related Stories