WebAnything: A Fully Automated Framework for Scalable Web Agent Environments — Artificial Intelligence

Researchers have introduced WebAnything, a new automated framework that reconstructs executable and resettable website environments directly from real URLs. By eliminating the need for manual per-site engineering, the system has generated 244 website artifacts across 14 domains, supporting nearly 13,000 tasks and demonstrating superior fidelity compared to human-assisted baselines.

A new study published on arXiv introduces WebAnything, a fully automated framework designed to address the critical need for diverse, executable, and resettable website environments for training and evaluating web agents. While existing methods often rely on labor-intensive human-assisted mirrors or generative construction that can diverge from real sites, WebAnything reconstructs source-grounded environments directly from real URLs without requiring manual implementation or repair after initial setup.

The framework operates through a coupled Explorer-Builder system that collects browser-observable structures, resources, and interaction outcomes. An independent verifier then checks frozen candidates against source evidence and site contracts. This verification process guides the system to either perform local repairs on implementation errors or trigger targeted recapture of source data when evidence is insufficient. This cycle ensures that the reconstructed environments maintain high fidelity to the original websites while remaining reusable for repeated testing.

In terms of scale, the researchers have constructed 244 website artifacts spanning 14 application domains and 53 categories. These environments support a total of 12,984 executed and accepted tasks. When compared against WebHarbor, a human-assisted baseline, WebAnything achieved higher aggregate scores on three fidelity metrics. In a blind audit involving five matched sites and 14 historical states, the automated framework secured 12 wins and two ties with no losses.

The utility of these environments was further demonstrated through downstream model training. Fine-tuning the Qwen3.5-9B model on 5,856 interaction trajectories collected from the WebAnything environments resulted in significant performance gains. The model saw a 9.9 percentage point improvement in REAL success rates, increases of 6.25 and 6.32 percentage points in Odysseys-Easy Rubric Avg and Task Avg respectively, and a 13.33 percentage point boost in WebArena-Verified Overall success under a fixed harness. The code for the project is available on GitHub.

Sources: arXiv cs.AI.