Browserenv

Browserbase's BrowserEnv, built with Prime Intellect, provides an environment where browser agents can practice tasks on real websites. A training run opens browser sessions, lets the model act and uses a verifier to judge completion. Those judgments supply feedback for reinforcement learning, the process of adjusting behavior through rewarded attempts.

Browserbase supplies the browsers and website access, while Prime Intellect supplies the training platform, so researchers need mainly a dataset of tasks. The article describes an end-to-end run fine-tuning an open vision-language model on WebVoyager tasks in a version of the environment without anti-bot walls, and says the model improved on multi-step tasks, but it supplies no numerical success-rate comparison. Its concrete contribution is an operational route for training and evaluating agents through website interaction.