We just made Astra's computer use 2.5x faster with Stagehand. Astra works by executing code against the a11y tree, we built a translator that turns Astra's playwright commands into Stagehand. Astra chose to batch the entire logo into one command and one-shotted it.

See how it works:

@Stagehanddev Would it be interesting to explore adding code mode to webmcp? Its a bit early as not enough websites covers it but would be super efficient btw we created a webmcp benchmark and just added Astra computer use vs webmcp, really interesting results

@Stagehanddev stagehand does support webmcp already, we do recommend webmcp if it exists, but we're very far from that for a majority of the internet

@Stagehanddev holy

@Stagehanddev

@Stagehanddev why not write in stagehand directly? does astra not style transfer properly?

@Stagehanddev models atm are post-trained on playwright, we've experimented with skills, prompt optimization etc but realized we should just get out of the way and let the model write playwright (it's already good at this) and then port it to stagehand (so it's faster and more token efficient)

@Stagehanddev Did the model just need to think longer? Or was it doing obviously sub par stuff?

@Stagehanddev same thinking level, what's happening in the video is the model is just choosing to batch commands together in stagehand which is why it was able to one shot the painting the playwright version can't batch and ends up doing the task sequentially

@Stagehanddev I’m asking what the failure case is when it is given stagehand directly.

@Stagehanddev ah lol, yeah it was just generating malformed code and not able to thinking multiple steps in advance

@Stagehanddev Why even need stagehand instead of using astra directly?

@Stagehanddev this is using astra with the codex harness too, it just literally translates the playwright that the model generates into stagehand so that it's faster. no cost whatsover just pure performance improvements

@Stagehanddev 2.5x faster astra computer use through stagehand is the kind of receipt people screenshot

@Stagehanddev bruh

@Stagehanddev one-shotting is all you need

@Stagehanddev A11y-tree execution plus generated Playwright is a compelling split: the model handles intent and batching, while Stagehand provides deterministic plumbing. The production test is replay fidelity after DOM drift, retries, and partial side effects—not the one-shot logo demo.

@Stagehanddev Batching only holds while the a11y tree is stable - one virtualized list or a late-mounting portal and step 3 targets a node that already moved. Do you re-snapshot inside a batch, or commit the whole thing blind?

@Stagehanddev the speedup is fun, but the real artifact is the translator between intent and the accessible tree. i keep looking for the awkward handoff where a fast run still needs a human to take the mouse.

@Stagehanddev 2.5x speed is a great upgrade