The AI cost of JavaScript frameworks

my experience is that one-shotting a simple front end works great. Then you want to add features and move things around and each turn starts getting slower and the single page with JS and CSS and HTML starts to turn into a 3000 line monster and your context window starts to creak under the strain of even simple changes, because the LLM has to keep more and more of the application in its head just to make relatively small changes without breaking something else. Either you should have started with some kind of framework with hygienic rules on how to separate concerns or at minimum a good set of coding standards that force the LLM to start building clean components from scratch. The benefit isn't necessarily cleaner code for its own sake, it's that the LLM can work on one part of the application without having to understand the whole thing every time. That being said, once you have a 3000 line prototype it's not that complex to simply ask the LLM to refactor it into a set of orthogonal components. As long as it was building unit tests using playwright or something this can turn your slop into framework code cleanly. The interesting question is whether you saved anything by putting off that structure in the first place.

I run into this problem turning claude artifacts into things I might want to maintain longer term.

The point of using a framework is to reduce mental load at expense of customizability. The moment you need to change something in the framework - this mental model falls apart.

Obvious solutions are to vendor your dependencies (compile the framework from owned source) or to drop to the lower level of abstraction.

LLMs should have no problem using browser/node APIs directly since they're very well used, well documented, and not changing as much as higher level software does.

HTMX is interesting.

I have most of my apps in it without much trouble.

One of my recent ones had a very complicated form and state management.

AI really struggled with HTMX in that use case.

Like the form might work but the page would take 30s to load. Or some state would randomly drop or a toggle wouldn't work after changing a dependency.

If I rewrote that in react it would have one shot it without a doubt. Some things are just better models for different problems.

For the record I eventually got enough tests in place to keep the agent honest and the HTMX version is now a lot faster anything I could have got out of react but yeah, it was an adventure.

This benchmark is too simplified to be meaningful. The choice of a framework is rarely based on speed alone, it factors in other elements like the user's domain knowledge, long term maintainability and the framework's track record itself w.r.t security and development of features.

In this regard alone, it will make a lot of sense to spend extra $2 and 2 mins to trade off for a long-term maintainable codebase and code you can reason about than just something made quickly just to get it deployed fast.

In my personal use case, that would be Svelte 5 for me and I'm sure it will be something else for everyone.

This is buried at the end.

> My goal here is not to write the end-all-be-all benchmark for “I’m a vibe-coder, which framework should I choose?” The point of this post is to demonstrate that it’s possible in the era of AI coding to still make decisions based on evidence rather than vibes.

But it’s obvious that the article will be understood as a useful benchmark, so author probably should have addressed that point earlier and more thoroughly.

Taking the authors pov for a moment, they just showed you how to put all your criticisms to work in making a newer better benchmark :)

Astro comes out on top with the best time, #2 lowest spend and zero client side JS. I know what I'm using for my next project.

Solid work, but it would be really cool if we can also test Pokemon motions and how the game responds in real-time. Is an agent able to think about reactions, directional movements and attack patterns of each Pokemon is where the real intelligence test lies.

Traditional languages may have their limitations, but [C++, Rust, JavaScript, TypeScript etc] can still do pretty well.

> in the agentic era, when developers are no longer typing code by hand

I hope I'll be the last person on Earth writing code by hand.

Java script frameworks are just extra boiler plate that isn't needed in the age of ai-coding.

Server side page generation with a sprinkle of js will give you the same results with GBS less of packages and comple time

Interesting, I’m just sad svelte didnt behave better because of the LLM reading migration guide on every answer

Interesting article. Kind of seems like you'd want a more complex task, there's not much differentiation here.

I mean it just seems like even in the agentic era, you should know when to use the right tool for the right job. The site shown seems perfect for htmx style websites, so idk why you would use an SPA. Still very cool experiment, appreciate the data and human writing!

Angular sucks, I always knew it, and now we have some data that it’s broken even for agents. I’m not surprised

Nice to see Nuxt holding its own. Astro or Nuxt are my gotos for client work and personal projects.

Glad to see Astro doing well here. It's by far my favorite way to build sites of all kinds, both content rich MPAs, and react heavy web apps.

It is fairly obvious that it is all going to be some kind of React/JSX with tailwind.

where is jquery? :(

htmx for the win, as always

Just use plain HTML for God's sake. HN demonstrates very well that you can do plenty without some Javascript monstrosity. At the very least, make your your websites degrades gracefully when JS is turned off.

Unfortunately the author is not only correct about front end development vibes, but due to this state of affairs, I ceased to recognize browser based UIs as being software development at all. Too much browser tech is broken to call it software. As for agents, good for them!

[flagged]

[flagged]