Glonce

WebAnything builds 244 reusable web environments from real URLs for agent training

researchThe Glonce editorsHugging Face🔥57Glonce rating57Impact45Novelty60Relevance95Surprise25Players10Credibility75
WebAnything builds 244 reusable web environments from real URLs for agent training

WebAnything is a framework, described in a paper listed on Hugging Face, that reconstructs real websites as reusable environments for training and evaluating web agents. It starts from real URLs and, according to the paper, needs no per-site manual implementation or repair after initial setup and authorization.

The paper frames the problem this way. Web agents need diverse, executable and resettable website environments so that training and evaluation can be reproduced. The authors say existing routes each fall short: human-assisted mirrors demand per-site engineering, generative construction can diverge from the real websites it is meant to match, and capture-and-replay preserves observations but limits interaction coverage to what was captured and what the replay can do.

WebAnything has two main parts. An Explorer-Builder collects browser-observable structures, resources and interaction outcomes. An independent verifier then checks frozen candidates and guides either local repair or targeted recapture from the source.

The authors built 244 website artifacts across 14 application domains and 53 categories. These support 12,984 executed and accepted tasks. Against the human-assisted WebHarbor, WebAnything achieves higher aggregate scores on three fidelity metrics, and it recorded 12 wins, two ties and no losses in a blind audit.

The authors also tested the environments as training data. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improved REAL success by 9.9 percentage points. On Odysseys-Easy, Rubric Avg rose by 6.25 points and Task Avg by 6.32 points. WebArena-Verified Overall success rose by 13.33 percentage points under the fixed harness. All of these are percentage-point gains, not relative ones. The code is on GitHub under the china-qijizhifeng organisation (Web-Anything).

Key facts

  • WebAnything is a fully automated framework that reconstructs source-grounded, reusable web environments from real URLs, with no per-site manual implementation or repair after initial setup and authorization.
  • The authors built 244 website artifacts across 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks.
  • Against human-assisted WebHarbor, it scores higher on three fidelity metrics and goes 12 wins, two ties, no losses in a blind audit.
  • Fine-tuning Qwen3.5-9B on 5,856 trajectories raised REAL success by 9.9 percentage points and WebArena-Verified Overall success by 13.33 percentage points under the fixed harness.
  • On Odysseys-Easy, Rubric Avg and Task Avg improved by 6.25 and 6.32 percentage points; code is on GitHub.

Why it matters

Training and testing web agents needs websites that can be run, reset and repeated. The authors argue that today's options each carry a cost: human-built mirrors need engineering for every site, generated sites can drift from the real ones, and capture-and-replay limits how much an agent can do. WebAnything aims to remove the per-site manual work, which is what would let the number of environments grow.

Who it affects

Teams that train or evaluate web agents, and researchers who build benchmarks for them, are the direct audience. The paper's fine-tuning experiments use Qwen3.5-9B, an open model size that smaller groups can work with.

How to use it

The code is linked at github.com/china-qijizhifeng/Web-Anything. The framework starts from real URLs and, by the authors' account, needs only initial setup and authorization before it runs without per-site manual implementation or repair. Its output is website artifacts that agents can interact with and that can be reset, so they can serve both as training environments and as evaluation targets.

How solid is it

This is a preprint with results reported by the authors themselves. The blind audit against WebHarbor (12 wins, two ties, no losses) and the benchmark gains are as stated in the abstract. The abstract does not name the three fidelity metrics, say who ran the blind audit, or give baseline scores for Qwen3.5-9B, so the absolute levels before and after fine-tuning are not visible from it.

Risks and caveats

The gains are percentage points, and without baseline scores their weight is hard to judge. The abstract does not say whether the fine-tuning gains were compared against a model trained on WebHarbor or other environments. It also does not give the number of tasks or sites used in each benchmark evaluation, or the time and cost of building a site. The claim of no per-site manual work applies after initial setup and authorization.