Introducing our Browser Agent Evals.
We benchmark frontier and open models using various harnesses on computer-use tasks.
Our evals are fully open source and reproducible via our Evals CLI.
Your agent pauses/ crashes mid task. Can a new process pick up where it left off?
Stagehand 4.4 adds browser.disconnect() so an agent can detach and leave the browser running and a new agent can reattach to the same browser and can complete the task.
This unlocks use cases like human handoffs, scheduled resumes, agent crashes.
In the clip, Pi asks for approval, waits 30 minutes, then reconnects and finishes the same checkout.
Fable 5.2 scores 92% on HardBench, our curated benchmark of difficult computer-use tasks.
It beats the previous champion, Opus 5.5, in accuracy, but costs 2.5x more and is 13.76% slower on average.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
The @LangChain team evaluated 7 different Browser Tools to power Managed Deep Agents.
It's no surprise that Stagehand was simultaneously the fastest, cheapest, and most accurate. Read the full report to learn how.
Browserbase is now available as a one-click install in the Antigravity Plugin marketplace.
Navigate sites, fill out forms, and take screenshots with a verified browser that doesn't get blocked.
Your agents are only as good as their context, and the most valuable context is often locked behind a login.
Introducing Authenticated Fetch, 1 request to instantly extract data from logged-in sites.
It’s 3x faster than a full browser session, and available on all plans today.
To do real work on your behalf, agents need to sign-in securely.
We’re excited to partner with 1Password so your agents can log in without ever seeing your passwords.
Agents request a login, you approve it in 1Password, then credentials are sent directly to the browser.
You can extract typed data from any website using @pydantic and Stagehand.
Using Pydantic's enforced typing and Stagehand's structured extract, you can build robust and self-healing web scraping pipelines.
Read our full blog post below.
New ship on https://t.co/cER97mCe5x - record mode!
-> fire up a browser session
-> show your agent what to do, click-by-click
-> hit export to download your finished workflow
you get a SKILL.md your agent can follow, plus scripts for @Stagehanddev, Playwright, or Puppeteer.
Browserbase Search is now powered by Parallel (@p0) as the default search provider.
Parallel Fast returns sub-second results when speed matters. Parallel Advanced goes deeper for multi-step research when quality matters most.
Gemini 4 Argon is live in Stagehand.
It outperforms Fable 5.1 on computer use tasks with @eve and @pidotdev harnesses, while being roughly around the same cost/speed.
You can try the model in our playground at https://t.co/7bFLxRjpBE.
We're sponsoring the AI Assistant meetup on Monday October 19th @ 7:30pm.
Come meet other builders and hear from teams building the infra powering the fastest growing assistants on the market.
Link to RSVP below.
Coolest new feature in Stagehand - we can now automatically detect & invoke WebMCP tools on the fly.
This unlocks interacting with all @stripe checkout pages without repeatedly polling for new tools.
Give it a try on 4.2!