Introducing: Browser Use 1.0.0: CDP is all you need 👀 We deleted 90% of the old library. Predefined state and action space make it worse. Now the agent just writes Python REPL on raw CDP. Simplicity wins. > same models, up to 50% better > 2-5x cheaper > 2-3x faster > SOTA web agent From now on we ship fast. Please pin with ==

Ran Muse Ember on hardest long horizon browser benchmark 👀 It's EXTREMELY cheap, pareto frontier. Tasks with 100s of browser steps cost CENTS, not USD > 63.9/100, same as GPT-6 Luna medium > 3 cents a task, 2.4x cheaper > best score under 5 cents congrats @alexandr_wang, cheap models that can actually drive a browser are huge

Your browser agent works in a demo. How does it handle 200 web tasks? Browser Use Benchmark is a public repository of browser-agent evaluations for builders testing web automation. It helps you move beyond spot checks by running BU Bench V2.1 tasks against weighted scoring rubrics, with saved evidence to inspect what happened. Key features: • 200-task evaluation – the default runner loads the full V2.1 task set • Partial-credit scoring – weighted rubrics capture more than a pass/fail result • Inspectable evidence – saved traces, screenshots, and task outputs help you investigate failures • Cloud or local browsers – run with Browser Use Cloud or your own Chrome/Chromium • Targeted test runs – select specific task IDs and adjust concurrency to fit your API limits Free public GitHub repo. Running evaluations requires your own API keys. Compare scores only on the same task set and revision. Link in the reply 👇

Subagents are the biggest cause of AI failures In most harnesses: 1. Main agent writes a summary of what it knows 2. Subagent starts from that summary only, and makes assumptions 3. Subagent works and sends back a new summary of what it did Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of So I tried something new: checkpoint agents The initial agent has complete context of what needs to be done. And that input is already cached Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix) I measured it on a 137k-token session with Opus 5.5: → Main agent does it on its own: 67 s, $0.38, context +8-17k → Checkpoint: 117 s, $0.68, context +2k → Classic subagent: 123 s, $0.57, context +4-5k But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers Don't design harnesses based on our assumptions of human behavior

wabi 🤝 @Shopify: your group chat just became your personal shopper. the camping trip. mom’s birthday. the themed bachelorette party. it all starts in the group chat. now you can shop there, too. just ask wabi to “find us a tent and sleeping bags.” it already knows the plan, everyone's tastes, budgets, and what you need, so recommendations are tailored to your group. with Shopify, wabi can search millions of merchants, compare products, check live prices and availability, and find discounts that actually work. once everyone agrees, check out securely with Shop Pay using your saved details, and track your orders right in the chat. find it. compare it. buy it. track it. all together. all in one place.

We don't care about concurrency, our infra is scalable af. If you want to run thousands of parallel browsers you can do that on PAYG plan, no commitment needed, just dm me. The more concurrency you have the more money we make. (other browser providers charge you for concurrency btw)

Exciting collaboration with the @browser_use team on Odysseys2, the next generation of our long-horizon web navigation benchmark: Data + Evals: Developed jointly with Browser Use, Odysseys2 contains 100 challenging tasks grounded in real user requests and executed on the live internet. Unlike the original Odysseys, which drew on user browsing histories, Odysseys2 reflects the distribution of real-world user requests collected by Browser Use, providing a more realistic evaluation of web agents.