WebAnything: From the Wild Web to Scalable Agent Environments

WebAnything. From the Wild Web to Scalable Agent Environments. By Tianci Xie and six others, from Zhejiang University, Nanyang Technological University, and one other institution. Published on the twenty-seventh of October.

Web agents need website environments to train and be evaluated in, and for that work to be reproducible those environments have to be diverse, executable, and resettable. Each existing route falls short somewhere. Human-assisted mirrors demand engineering for every site. Generative construction can drift away from the real website it was meant to match. Capture-and-replay preserves what was observed, but interaction coverage is limited by what was captured and what the replay can do.

WebAnything is a fully automated framework that reconstructs Web environments from real URLs. The environments are source-grounded, meaning tied to the real site, and reusable. After initial setup and authorization, no site needs manual implementation or manual repair.

Two roles do the work. An Explorer-Builder collects what a browser can observe of the site: its structures, its resources, and the outcomes of interacting with it. Then an independent verifier checks the frozen candidate environment. Where it finds problems, it guides either a local repair or a targeted recapture from the source site.

With it, the authors built two hundred and forty-four website artifacts across fourteen application domains and fifty-three categories, supporting twelve thousand nine hundred and eighty-four tasks that were executed and accepted.

Against the human-assisted WebHarbor, WebAnything achieves higher aggregate scores on three fidelity metrics. In a blind audit it records twelve wins, two ties, and no losses.

The environments were also used for training. Fine-tuning Qwen3.5-9B on five thousand eight hundred and fifty-six interaction trajectories improves success on REAL by nine point nine percentage points. On Odysseys-Easy, the Rubric Avg rises by six point two five points and the Task Avg by six point three two points. And on WebArena-Verified, Overall success improves by thirteen point three three percentage points, under the fixed harness.

What WebAnything changes is where the human effort sits. Instead of engineering each site by hand, the effort goes into setup once, and after that an explorer and a verifier between them build, check and repair each environment from the real page.