Excited to share more of our latest work with Browserbase and Prime Intellect.
Until today, training computer-use models on real browser tasks was only possible if you were a frontier lab. The infrastructure to train browser agents didn't exist. You had to build it yourself: session management, environment resets, scalable browser access, and a way to survive anti-bot systems that treat your training agent like a scraper. Sessions die mid-trajectory and you never accumulate the data you need. We heard this repeatedly from customers. Today, Browserbase and Prime Intellect are launching BrowserEnv to change that. BrowserEnv is a reinforcement learning environment built for browser agents. Browserbase handles the browser infrastructure and real website access, while Prime Intellect handles the training platform. All you need to do is bring a dataset of tasks. Microsoft used Browserbase to train and evaluate Fara-7B: a workflow that required reliable access to real websites and scalable environments for RL. We also used it internally to fine-tune Qwen3-VL-8B on WebVoyager and saw meaningful improvement on benchmark tasks. For product teams, you don't have to use a general-purpose model and hope it navigates your product correctly. You can train something specific to your workflows on real browser environments, without a large infrastructure team and without building from scratch. The barrier to entry just dropped significantly. Check out -> https://lnkd.in/gcGfVZ2U