Training web agents to learn from live websites, not just recorded demos

Building capable web agents usually requires huge collections of hand-curated website interaction recordings—expensive and incomplete. OpenWebRL trains visual agents directly on live websites using reinforcement learning instead, letting them learn from real-world feedback. The framework achieved 67% success on Online-Mind2Web and 64% on DeepShop with minimal supervised data, matching OpenAI's and Google's proprietary systems. Code and models will be released.