we reverse engineered astra's computer use and made it 2x faster.

how? well you should start by understanding how Astra's computer use works. i did a deep dive into the source code and wrote an article about it so you don't have to.

tl;dr it uses code mode + the a11y tree to be more accurate and faster. Astra defaults to Playwright, which it's post-trained on, so it's really good at writing and executing it.

we just turn playwright into stagehand code (our sdk for browser agents) which brings a speed up in action execution time with "batching". the models are very good at predicting next step, but now they can predict the next 5, 10, even 20 steps. batching lets us group together computer use steps and dispatch them all at once.

in the end, it's faster and in many cases qualitatively better since it's thinking pattern is different, like how it chose to paint Starry night here in the demo.

here's how we did it:

So isn't this browser use and not computer use?

cool so the efficiency gain stems mostly from stagehand being faster than playwright in operating the browser?

being faster in both action execution, but also the batching. doing sequential actions is pretty but not as efficient as bundling them together and dispatching. it's even more obvious when you use a remote browser, cdp rtt is no joke

dispatching i.e. executing in parallel?

more like sending a bunch of sequential calls but really fast and back to back to back s.t. it seems like it's in parallel sequential calls would flow: > figure out what to do > do it > wait for response > repeat whereas batching: >what are the next 5 things to do > do them > wait for response >

got it so less round trips..really interesting stuff!

precisely!

I created a similar version of this using Rust, got a very nice results as well, if you wanna I can send for you guys to use/explore :)

Genius 😎

Whoa