Jeff Huang

i code designs and design code

eng @AdaptiveSec

[Jeff Huang @jeffzxh]

@sidpalas @toolmantim @AmpCode i see - so intended for non monorepo setups i guess. makes sense

[Jeff Huang @jeffzxh]

the delta between frontier models is so small now, parallelization is a bigger multiplier

maybe that’s still called the harness, who cares

what’s for sure is i care way more about running 20 dumb agents at once than one smart agent alone on my laptop. this was true with cursor cloud agents, orbs have improved that experience even more

[Jeff Huang @jeffzxh]

@beyang @thorstenball @toolmantim https://t.co/YBFSqDIoRc i wrote that before seeing this

ignore the noise, Amp consistently sets the bar for others to follow. the average dev is still learning to use claude code (not to say prime is an average dev)

[Jeff Huang @jeffzxh]

as a thought experiment i rewrote our BE API in typescript... it took a week, i was completely hands off, prompting my agent ~2-3x a day

A month ago i think this would've taken 3-5x as long. 2 key learnings:

[Jeff Huang @jeffzxh]

both these learnings are different ways to parallelize agents:

models are more than capable, it's now about how to drive them to solve multiple problems at once or large difficult problems in pieces, with minimal supervision

[Jeff Huang @jeffzxh]

as a thought experiment i rewrote our BE API in typescript... it took a week, i was completely hands off, prompting my agent ~2-3x a day

A month ago i think this would've taken 3-5x as long. 2 key learnings:

[Jeff Huang @jeffzxh]

2: local dev is dead

I haven't pulled main at all this month. I've been preaching @cursor_ai Cloud Agents for a while, and Amp orb's have re-solidified my opinion:

cloud agents lets you do multiple tasks at once, without local resource constraints.

cloud agents are a bigger unlock now than model performance. frontier models are pretty close in performance, but running a model on N problems at once is far more effective than running a model that is 1.1x better on 1 problem

the ability to run your entire stack in a cloud VM should be a top priority for every team, along with investing in harnesses that prioritize cloud

[Jeff Huang @jeffzxh]

the only thing stopping AI models from discovering zero-days everywhere are the constraints that frontier labs are imposing... i have to imagine that sooner rather than later that the agents disregard these guardrails altogether. this is only the 1st case

[Jeff Huang @jeffzxh]

insane. @OpenAI /@jfrog patched the exploit the model found to gain network access... but by then it was already used to train the next iteration of the model.

that knowledge was used gain admin control of the cluster and eventually exploit hugging face

https://t.co/oSdHRDQImY

[Jeff Huang @jeffzxh]

@github's uptime is all the advertising that cursor origin needs

[Jeff Huang @jeffzxh]

@sqs @AmpCode tiny request: denote between "root" level threads created from another thread.

now that my agents are spawning so many threads i have trouble finding the original top level thread to prompt more

or maybe it's a skill issue

[Jeff Huang @jeffzxh]

every time i try using claude code or codex it feels like i’m on a version 3 months outdated

@cursor_ai has had cloud agents for months

@AmpCode orbs, multiplayer, and agent <> agent execution are incredible

access to multiple models is 👑

i guess it shouldn’t be a surprise the best harnesses are still being built by the companies that specialize in them

[Jeff Huang @jeffzxh]

@MerrillLutsky @graphite amazing run! you quite literally took graphite to the moon

[Jeff Huang @jeffzxh]

@sqs @AmpCode @thorstenball sure, will dm you the report IDs!

[Jeff Huang @jeffzxh]

hahahah is this what they meant by "loop" engineering?? 🤣 @AmpCode @sqs @thorstenball

[Jeff Huang @jeffzxh]

@ahejlsberg @typescript woo! is there any recommendations for running side by side with TS 6 for projects that need the JS API, that supports pnpm?

[Jeff Huang @jeffzxh]

@shadcn @base_ui Exciting and the right direction

but it’s a shame Ark UI isn’t supported yet - it’s further ahead than both base and radix

[Jeff Huang @jeffzxh]

@GergelyOrosz CLAUDE.md tells us all we need to know about their stance on vendor lock in 🙄

[Jeff Huang @jeffzxh]

@kenneth_skovhus @linear @stylexjs wait Linear doesn’t have a design system?? 🤯

we chose Panda over StyleX 2 years ago, i wonder if we would have made the same choice today