WebAnything automates cloning of browsing environments, outperforms manual simulation and marks a leap in agent performance
The development of intelligent web agents faces a silent operational dilemma: language models can understand commands, but reliable training and evaluation require realistic browsing environments that are executable and resettable after each run. The options available to date have cornered developers into costly compromises; manual site simulation demands laborious engineering for each individual site, generative simulation quickly deviates from the web’s real logic, while logging and replay systems fail to provide broad interaction paths beyond what was recorded in the first session.
In a new research paper published by researcher Tianchi Xie and colleagues on the arXiv platform, the research team presents the WebAnything framework, a fully automated system for reconstructing real, reusable web environments from direct site URLs, without the need for manual coding or site-specific fixes after the initial setup and authorization stage.The framework overcomes the bottleneck of manual construction through a double-layered architecture that separates collection from field verification.
The system relies on two main engines: the first, a crawler-builder, gathers observable structures through the browser, resources, and programmatic interaction results, while the second acts as an independent auditor that examines frozen copies, directs localized fixes, or precisely re-captures targeted data from the source whenever a structural flaw is detected.
The framework produced 244 interactive site environments covering 14 application domains and 53 diverse categories, supporting 12,984 programmatically executed and accepted tasks. In a blind audit comparing the framework to the WebHarbor system, which relies on direct human engineering, WebAnything achieved superior overall results across three matching-accuracy metrics, recording 12 wins and two ties with no losses.
The true value of the interactive environments becomes evident when measuring their impact on model training; the researchers performed a precise fine-tuning of the open-weight model Qwen-3.5, which has nine billion parameters, using 5,856 interaction traces extracted via the framework.The results showed an immediate jump in model efficiency of 13.33 percentage points on the documented WebArena tests under the fixed-evaluation platform.Success rate on the Real benchmark also rose by 9.9 percentage points, and the Odyssey Easy scores for average evaluation metrics and tasks increased by 6.25 and 6.32 percentage points respectively, with the full project code made available.
This technological shift has a tangible impact on engineering and digital-transformation teams in the Gulf, Egypt, and the Levant; organizations that build automation agents for banking services, e-commerce platforms, or e-government face a real risk when testing models on live gateways and incur high costs in manually building experimental test environments. I see automation of gateway cloning as a solution that enables our development teams to generate safe, locally-aligned simulation environments at low cost, accelerating agent testing and qualification before deployment in production environments.
At the infrastructure-choice level, the recorded superiority of the nine-billion-parameter model demonstrates that a technical decision is no longer constrained by costly closed-cloud model-calling invoices for complex browsing tasks; an integrated, open model fine-tuned on precise interaction paths suffices to achieve reliable operational results within local data centers at low inference cost.