WebAnything: A Framework for Scalable Web Agent Training

Paper: WebAnything: From the Wild Web to Scalable Agent Environments

Authors: Tianci Xie, Zikun Cai, Chaoyun Zhang, Pinxue Guo, Bingxuan Li, Rui Zheng, Hang Yan

Executive Summary

The research introduces WebAnything, an innovative framework designed to create reusable web environments for training and evaluating web agents. Traditional methods often rely on manual engineering of specific sites or on generative techniques that may not accurately reflect real websites. This can lead to inefficiencies and inconsistencies in how web agents learn and perform.

WebAnything automates the process of reconstructing web environments from real URLs, eliminating the need for manual setup or ongoing repairs once the initial configuration is done. This is particularly beneficial for organizations looking to train web agents without the labor-intensive work typically required. The framework consists of two main components: the Explorer-Builder, which gathers observable data from browsers, and an independent verifier that assesses these data collections and aids in any necessary corrections or adjustments.

In their study, the researchers developed 244 unique web artifacts across 14 application domains and 53 categories, allowing for a broad range of tasks to be executed and evaluated. Specifically, WebAnything supported a total of 12,984 tasks, demonstrating its capability to handle extensive interaction scenarios.

When benchmarked against an existing tool, WebHarbor, WebAnything performed better across several metrics designed to assess fidelity. In a blind audit, WebAnything achieved 12 wins, two ties, and did not lose any comparisons. This suggests that it provides a more reliable platform for training web agents than previous methods.

Additionally, fine-tuning an AI model named Qwen3.5-9B with data from 5,856 interaction trajectories led to notable improvements in performance. The results indicated increases in success rates: 9.9 percentage points in overall task success, 6.25 and 6.32 percentage points in two specific evaluation metrics, and a 13.33 percentage point boost in a verified success metric.

These findings indicate that WebAnything could significantly enhance the efficiency and effectiveness of developing and training web agents. Organizations may find this framework particularly useful in sectors where web interactions are critical, such as e-commerce, customer service, and digital marketing. By leveraging WebAnything, businesses could potentially reduce the time and resources spent on manual site engineering, allowing for more rapid and reliable development of web-based AI applications.

Academic 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 URLs without per-site manual implementation or repair after initial setup and authorization. An Explorer-Builder collects browser-observable structures, resources, and interaction outcomes; an independent verifier checks frozen candidates and guides local repair or targeted source recapture. We construct 244 website artifacts across 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks. Compared with human-assisted WebHarbor, WebAnything achieves higher aggregate scores on three fidelity metrics, with 12 wins, two ties, and no losses in a blind audit. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improves REAL success by 9.9 percentage points, Odysseys-Easy Rubric Avg and Task Avg by 6.25 and 6.32 percentage points, and WebArena-Verified Overall success by 13.33 percentage points under the fixed harness. Code: https://github.com/china-qijizhifeng/Web-Anything