stefan brunthaler
Andy Wingoa colleague of mine, on a whim, told a coding agent to make a js implementation that passes all of test262. such a thing does not exist in the wild, despite hundreds of millions of investment in JS. the results are much better than i had expected https://p.ocmatos.com/blog/jsse-a-javascript-engine-built-by-an-agent.html
Philn@wingo In this becoming a trend? https://iangrunert.com/2026/03/22/writing-a-js-engine-using-claude-code
@philn omg it is in the water
Edward Hervey@wingo as long as you don't have to maintain the code it produced 😅
@bilboed he is treating this as a write-only experiment 😅
Henri Sivonen@wingo My first check is the thing that I’ve most recently worked on in SpiderMonkey, and the agent-written code is holding the collator wrong: https://github.com/pmatos/jsse/blob/c490bc3fee8c4c8e6b2452f4acacb60d84b02a39/src/interpreter/builtins/intl/collator.rs#L90
(The collator normalizes internally. No point normalizing before calling into the collator. I’m curious if this is the agent doing bogus things or the prompts saying bogus things.)
jsse/src/interpreter/builtins/intl/collator.rs at c490bc3fee8c4c8e6b2452f4acacb60d84b02a39 · pmatos/jsse
An agent-coded JS engine in Rust. I didn't touch a single line of code here. Not one. This repo is a write-only data store. I didn't even create this repo by hand -- my agent did that. - pm...
@wingo Oh, it’s even worse: it instantiates a new collator for each comparison.
On one hand, it’s amazing that a robot can make code pass tests. On the other hand, if this is the direction of the profession, there’s going to be a lot of bad code.
@hsivonen it's a very interesting experiment and i have no idea what to make of it 😅
sayrer@wingo @hsivonen take it easy! you can always use with prompts like "find bugs in this code" or "make it faster".
I didn't file these yet, because I was being annoying in this exact way, but here it is:
https://gist.github.com/sayrer/9139640efd3496357447e1ca5a614364
