WebHarbor suite provides real website environments for GUI agents
Original title: WebHarbor: Docking Real Websites for Evolving GUI Agent Environments
WebHarbor offers 153 executable dockerized environments with 3518 tasks for evolving GUI agent training and evaluation.
WebHarbor reconstructs visual interfaces, content, and functional workflows of real-world web services under deterministic control. Environments evolve through task-driven human-agent collaboration where coding agents implement changes and humans audit fidelity. A Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively across four benchmarks.
- 153 executable dockerized environments
- 3518 multi-step and post-authentication tasks
- Evolves via human-agent collaboration
- High-fidelity for RL training
From the original article
1 min read at arxiv.orgGUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity.
By Zhaoyang Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.