WebAnything Automates Real‑Web Environments for AI Agents
Training and evaluating web‑based AI agents demands environments that are diverse, executable, and can be reset for reproducible experiments. Traditional solutions rely on human‑assisted mirrors, which require per‑site engineering, or on generative methods that may drift from actual sites. Capture‑and‑replay preserves observations but limits interaction to what was recorded. WebAnything addresses these gaps with a fully automated pipeline that builds source‑grounded, reusable web environments directly from live URLs, eliminating the need for manual site‑specific code after initial authorization.
The system’s Explorer‑Builder crawls browser‑observable structures, resources and interaction outcomes, while a separate verifier freezes candidates and triggers local repairs or targeted recaptures when needed. Using this workflow, the authors generated 244 website artifacts spanning 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks. In a blind audit against the human‑assisted WebHarbor benchmark, WebAnything achieved higher aggregate scores on three fidelity metrics, recording 12 wins, two ties and no losses.
Further gains were demonstrated by fine‑tuning the Qwen3.5‑9B model on 5,856 interaction trajectories, which lifted REAL success by 9.9 percentage points, improved the Odysseys‑Easy Rubric average and task average by 6.25 and 6.32 points respectively, and raised the WebArena‑Verified overall success by 13.33 points under a fixed harness. (Opinion: The automation reduces labor‑intensive site engineering, making large‑scale agent training more scalable and reliable, which should interest AI researchers, developers of web‑automation tools, and laboratories integrating AI into LIMS workflows.)
Why it matters
Automated, high‑fidelity web environments lower the cost of building and testing AI agents at scale.
FAQ
How many website artifacts does WebAnything create?
It constructs 244 website artifacts across 14 application domains and 53 categories.
What performance improvements were observed after fine‑tuning Qwen3.5‑9B?
Fine‑tuning on 5,856 interaction trajectories improved REAL success by 9.9 percentage points, Odysseys‑Easy Rubric Avg and Task Avg by 6.25 and 6.32 points, and WebArena‑Verified Overall success by 13.33 points under the fixed harness.
Source: arXiv - AI Research Papers. See the original source for full details.
Scan with your phone
Scan with your phone