WebAnything：从真实网页自动构建可扩展智能体环境
原标题：WebAnything: From the Wild Web to Scalable Agent Environments
AI 摘要
研究团队提出 WebAnything 框架，可从真实 URL 全自动重建可执行、可重置的网站环境，无需逐站人工实现。该框架通过 Explorer–Builder 采集浏览器可观测结构与交互结果，并由独立验证器检查冻结候选、指导局部修复或定向重采集。团队构建了覆盖 14 个领域、53 个类别的 244 个网站工件，支持 12,984 个已执行任务；在盲审中对比 WebHarbor 取得 12 胜 2 平 0 负，且用 5,856 条交互轨迹微调 Qwen3.5-9B 后，REAL 成功率提升 9.9 个百分点。
正文节选
Abstract Web agents require diverse, executable, and resettable website environments for reproducible training and evaluation. Human-assisted mirrors demand per-site engineering, while generative construction can diverge from specified real websites. Capture-and-replay preserves observations, with interaction coverage constrained by captured content and replay capabilities. We present WebAnything, a fully automated framework that reconstructs source-grounded, reusable Web environments from real