Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Scaling long-running autonomous coding

cursor.com

Post reply on HN
Print view View original

41–50 of 203 posts

Re: Scaling long-running autonomous coding

#41

Post by Snuggly73 »

micimize wrote: post #38
embedding-shape wrote: post #26

Earlier quoted context omitted.

> it's a mountain of inscrutable agent output that manages to compile But is this actually true? They don't say that as far as I can tell, and it also doesn't compile for me nor their own CI it seems.

Hah I don't know actually! I was assuming it must if they were able to get that screenshot video.

error: could not compile `fastrender` (lib) due to 34 previous errors; 94 warnings emitted

I guess probably at some point, something compiled, but cba to try to find that commit. I guess they should've left it in a better state before doing that blog post.

Re: Scaling long-running autonomous coding

#42

Post by rvz »

sashank_1509 wrote: post #36
embedding-shape wrote: post #26

Earlier quoted context omitted.

> it's a mountain of inscrutable agent output that manages to compile But is this actually true? They don't say that as far as I can tell, and it also doesn't compile for me nor their own CI it seems.

Oh it doesn’t compile? that’s very revealing

Some people just believe anything said on X these days. No timeline from start to finish, just "trust me bro".

If you can't reproduce or compile the experiment then it really doesn't work at all and nothing but a hype piece.

Re: Scaling long-running autonomous coding

#43

Post by jaggederest »

Snuggly73 wrote: post #41
micimize wrote: post #38

Earlier quoted context omitted.

Hah I don't know actually! I was assuming it must if they were able to get that screenshot video.

error: could not compile `fastrender` (lib) due to 34 previous errors; 94 warnings emitted I guess probably at some point, something compiled, but cba to try to find that commit. I guess they should've left it in a better state before doing that blog post.

I find it very interesting the degree to which coding agents completely ignore warnings. When I program I generally target warning-free code, and even with significant effort in prompting, I haven't found a model that treats warnings as errors, and they almost all love the "ignore this warning" pragmas or comments over actually fixing them.

Re: Scaling long-running autonomous coding

#44

Post by rvz »

mk599 wrote: post #16

Define "from scratch" in "building a web browser from scratch". This thing has over 100 crates as dependencies... To implement css layouting, it uses Taffy, a crate used by existing browser implementations...

When I see hundreds of crates being used in a project, I have to just scratch my head and ask: What the f___?

If one vulnerability exists in those crates well, thats that.

Re: Scaling long-running autonomous coding

#45

Post by xenni »

bob1029 wrote: post #22
simonw wrote: post #3

"To test this system, we pointed it at an ambitious goal: building a web browser from scratch." I shared my LLM predictions last week, and one of them was that by 2029 "Someone will build a new browser using mainly AI-assisted coding and it won’t even be a surprise" https://simonwillison.net/2026/Jan/8/llm-predictions-for-202... and https://www.youtube.com/watch?v=lVDhQMiAbR8&t=3913s This project from Cursor is the s…

The goal I am currently using for long horizon coding experiments is implementation of a PDF rasterizer given an ISO32000 specification document.

We're almost there, I've been working on something similar using a markdown'd version of the ISO32000 spec

Re: Scaling long-running autonomous coding

#46

Post by flyinglizard »

tired_and_awake wrote: post #28

The moment all code is interacted with through agents I cease to care about code quality. The only thing that matters is the quality of the product, cost of maintenance etc. exactly the thing we measure software development orgs against. It could be handy to have these projects deployed to demonstrate their utility and efficacy? Looking at PRs of agents feels a wrong headed, like who cares if agents code is hard to r…

You could look at agents as meta-compilers, the problem is that unlike real compilers they aren't verified in any way (neither formally or informally), in fact you never know which particular agent you're running against when you're asking for something; and unlike compilers, you don't just throw away everything and start afresh on each run. I don't think you could test a reasonably complex system to a degree where it really wouldn't matter what runs underneath, and as you're going to (probably) use other agents to write THOSE tests, what makes you certain they offer real coverage? It's turtles all the way down.

Re: Scaling long-running autonomous coding

#47

Post by jaggederest »

simonw wrote: post #5
jphoward wrote: post #2

The browser it built, obviously the context window of the entire project is huge. They mention loads of parallel agents in the blog post, so I guess each agent is given a module to work on, and some tests? And then a 'manager' agent plugs this in without reading the code? Otherwise I can't see how, even with ChatGPT 5.2/Gemini 3, you could do this otherwise? In retrospect it seems an obvious approach and akin to how…

GPT-5.2-Codex has a 400,000 token window. Claude 4.5 Opus is half of that, 200,000 tokens. It turns out to matter a whole lot less than you would expect. Coding Agents are really good at using grep and writing out plans to files, which means they can operate successfully against way more code than fits in their context at a single time.

The other issue with "a huge token window" is that if you fill it, it seems like relevance for any specific part of the window is diminished - which makes it hard to override default model behavior.

Interestingly, recently it seems to me like codex is actually compressing early and often so that it stays in the smarter-feeling reasoning zone of the first 1/3rd of the window, which is a neat solution for this, albeit with the caveat of post-compression behavior differences cropping up more often.

Re: Scaling long-running autonomous coding

#48

Post by leptons »

simonw wrote: post #3

"To test this system, we pointed it at an ambitious goal: building a web browser from scratch." I shared my LLM predictions last week, and one of them was that by 2029 "Someone will build a new browser using mainly AI-assisted coding and it won’t even be a surprise" https://simonwillison.net/2026/Jan/8/llm-predictions-for-202... and https://www.youtube.com/watch?v=lVDhQMiAbR8&t=3913s This project from Cursor is the s…

Great, they can call it "artificial Internet Explorer", or aIE for short.

Re: Scaling long-running autonomous coding

#49

Post by rvz »

xmprt wrote: post #12
cheevly wrote: post #9

Earlier quoted context omitted.

2029? I have no idea why you would think this is so far off. More like Q2 2026.

You're either overestimating the capabilities of current AI models or underestimating the complexity of building a web browser. There are tons of tiny edge cases and standards to comply with where implementing one standard will break 3 others if not done carefully. AI can't do that right now.

It's most likely both.

> There are tons of tiny edge cases and standards to comply with where implementing one standard will break 3 others if not done carefully. AI can't do that right now.

Firstly the CI is completely broken on every commit, all tests have failed and its and looking closely at the code, it is exactly what you expect for unmaintainable slop.

Having more lines of code is not a good measure of robust software, especially if it does not work.

Re: Scaling long-running autonomous coding

#50

Post by tired_and_awake »

flyinglizard wrote: post #46
tired_and_awake wrote: post #28

The moment all code is interacted with through agents I cease to care about code quality. The only thing that matters is the quality of the product, cost of maintenance etc. exactly the thing we measure software development orgs against. It could be handy to have these projects deployed to demonstrate their utility and efficacy? Looking at PRs of agents feels a wrong headed, like who cares if agents code is hard to r…

You could look at agents as meta-compilers, the problem is that unlike real compilers they aren't verified in any way (neither formally or informally), in fact you never know which particular agent you're running against when you're asking for something; and unlike compilers, you don't just throw away everything and start afresh on each run. I don't think you could test a reasonably complex system to a degree where i…

Completely agree and great points. The conclusion of "agents are writing the tests" etc is where I'm at as well. More over the code quality itself is also an agentic problem, as is compile time, reliability, portability... Turtles all the way down as you say.

All code interactions all happen through agents.

I suppose the question is if the agents only produce Swiss cheese solutions at scale and there's no way to fill in those gaps (at scale). Then yeah fully agentic coding is probably a pipe dream.

On the other hand if you can stand up a code generation machine where it's watts + Gpus + time => software products. Then well... It's only a matter of time until app stores entirely disappear or get really weird. It's hard to fathom the change that's coming to our profession in this world.

Post reply on HN