We built Stagehand 2 years ago (24k stars and 4M monthly npm downloads) and recently fixed its biggest flaw: round-trip latency.
Every action performed requires a round trip between your script and the browser (short when running locally but increased when running in the cloud). We also saw multiple posts complaining about the eager token appetite of Playwright MCP.
For this reason, we rebuilt Stagehand from the ground up and shipped v4, where Stagehand controls the browser from an extension automatically loaded upon your browser startup.
Stagehand v4 comes with batch command support, dedicated token-efficient methods `act()` and `extract()`, and a brand new architecture making it 2x faster than Playwright and 80% more token efficient.
You can see for yourself by looking at our benchmarks, comparing its performance across a dozen models (frontier and open weights) and tools (Codex, Claude Code, and more): https://www.stagehand.dev/evals
Looks very useful, and I like the caching idea which I think makes it interesting for self-healing CI tests.
How does it determine when a cached act() fails and has to be re-evaluated by the LLM? And in particular, if the cache is saved in the cloud (Browserbase?), won't this lead to a lot of cache churn if used in CI pipelines where different versions of the site are running against the same cache?
Also, is there a technical reason why the cache couldn't just be a local file that's checked in along with the script but must be provided by Browserbase? If it was, devs could heal failing tests locally using LLM calls, while CI runs entirely deterministically.
Exactly, as Stagehand now runs inside the browser, you'll save on the round trip.
Also, enabling batch actions will further accelerate your test suite.
seems really cool! although, one question i have is why not keep agent() alongside the new primatives? it seems v4 removed agent() entirely rather than offering it with all of act(), observe(), extract().
We removed agent because so many great harnesses are available in the ecosystem.
Instead of keeping it, we decided to make Stagehand v4 better integrated with popular harnesses, both at the Coding Agent level (Codex, Claude Code) and frameworks level (Eve, Deep Agents, Mastra, etc)
It still uses CDP but communicates from an extension within the browser instead of a script running in a separate runtime or, worse, in a separate region.
A lot, especially for performance. That's why we built our own benchmarks that account for both the model and the harness.
For example, with Claude Opus 5, the accuracy gap can be up to 3% and performance up to 200ms, depending on whether you're using Deep Agents, Eve, or Fx.
I tried this myself - my conclusion was the overhead of parsing a screenshot, generating an action, and being limited to headful mode is much less efficient than reading accessibility trees & generating CDP commands.
basically, browser automation is a closer-to-the-metal abstraction than computer use, allows more flexibility, and ends up being much cheaper at scale!
We built Stagehand 2 years ago (24k stars and 4M monthly npm downloads) and recently fixed its biggest flaw: round-trip latency.
Every action performed requires a round trip between your script and the browser (short when running locally but increased when running in the cloud). We also saw multiple posts complaining about the eager token appetite of Playwright MCP.
For this reason, we rebuilt Stagehand from the ground up and shipped v4, where Stagehand controls the browser from an extension automatically loaded upon your browser startup.
Stagehand v4 comes with batch command support, dedicated token-efficient methods `act()` and `extract()`, and a brand new architecture making it 2x faster than Playwright and 80% more token efficient.
You can see for yourself by looking at our benchmarks, comparing its performance across a dozen models (frontier and open weights) and tools (Codex, Claude Code, and more): https://www.stagehand.dev/evals
Ask me anything!
Looks very useful, and I like the caching idea which I think makes it interesting for self-healing CI tests.
How does it determine when a cached act() fails and has to be re-evaluated by the LLM? And in particular, if the cache is saved in the cloud (Browserbase?), won't this lead to a lot of cache churn if used in CI pipelines where different versions of the site are running against the same cache?
Also, is there a technical reason why the cache couldn't just be a local file that's checked in along with the script but must be provided by Browserbase? If it was, devs could heal failing tests locally using LLM calls, while CI runs entirely deterministically.
If your needs are simpler, I created a tiny headless WebKit browser specifically for agents called Sleepy Hollow [0]
[0]: https://github.com/bensyverson/sleepyhollow
What would be the Linux equivalent?
I love the name.
Would you recommend to use this over agent-browser for general agent-based validation work? Any performance benefit?
So for an enterprise with a 1,000 test Playwright suite, does this basically mean ~2x faster CI times? That would be huge.
Exactly, as Stagehand now runs inside the browser, you'll save on the round trip. Also, enabling batch actions will further accelerate your test suite.
judging from their documentation looks like promised increase only achievable with their main product
seems really cool! although, one question i have is why not keep agent() alongside the new primatives? it seems v4 removed agent() entirely rather than offering it with all of act(), observe(), extract().
We removed agent because so many great harnesses are available in the ecosystem. Instead of keeping it, we decided to make Stagehand v4 better integrated with popular harnesses, both at the Coding Agent level (Codex, Claude Code) and frameworks level (Eve, Deep Agents, Mastra, etc)
what did you lose compared to CDP (e.g. cross-origin iframes, downloads running in envs where you can't load extensions)?
It still uses CDP but communicates from an extension within the browser instead of a script running in a separate runtime or, worse, in a separate region.
Do I need to pay for Browserbase to use this?
Nope, Stagehand is open-source and works with local browsers by default.
no matter how much faster you make.. playwright is playwright. Dead bot giveaway.
A lot of folks use Playwright for UI testing.
How much does the harness really matter for evals?
A lot, especially for performance. That's why we built our own benchmarks that account for both the model and the harness. For example, with Claude Opus 5, the accuracy gap can be up to 3% and performance up to 200ms, depending on whether you're using Deep Agents, Eve, or Fx.
More details here: https://www.stagehand.dev/evals
lowk why not just do astra computer use
I tried this myself - my conclusion was the overhead of parsing a screenshot, generating an action, and being limited to headful mode is much less efficient than reading accessibility trees & generating CDP commands.
basically, browser automation is a closer-to-the-metal abstraction than computer use, allows more flexibility, and ends up being much cheaper at scale!
Astra with computer use seems to burn a ton of tokens though. Stagehand seems more token efficient.