WebAnything Turns Real Websites into Reusable Agent Environments

Web agents need more than screenshots or a handful of manually coded demos. To learn how to search, click, fill forms, and complete multi-step tasks, they require environments that are executable, diverse, resettable, and reproducible. Existing approaches make that difficult. Human-assisted mirrors often require site-specific engineering, while generative replicas can drift away from the real website. Capture-and-replay preserves observed pages, but its interaction space is limited by what was captured and what the replay system can reproduce.

An automated reconstruction pipeline

WebAnything is designed to turn a real URL into a reusable web-agent environment without requiring a separate manual implementation for every site. The framework has two central components:

• Explorer-Builder collects browser-observable page structures, resources, and outcomes produced by user interactions, then uses them to construct an environment.
• An independent verifier examines frozen candidate environments. When it detects a mismatch, it can guide local repair or request targeted recapture from the source website.

The distinction matters because the goal is not merely visual imitation. A useful agent environment must also preserve executable actions and their consequences. According to the paper, after the initial setup and authorization, WebAnything does not require per-site manual implementation or post-construction repair by hand.

Reported scale and results

The authors built 244 website artifacts spanning 14 application domains and 53 categories. These environments supported 12,984 executed and accepted tasks. In a blind comparison with the human-assisted WebHarbor system, WebAnything achieved higher aggregate scores on three fidelity metrics, with 12 wins, two ties, and no losses.

The environments were also used as a source of interaction data. Fine-tuning Qwen3.5-9B on 5,856 trajectories increased REAL success by 9.9 percentage points. Under the fixed evaluation harness reported in the paper, Odysseys-Easy Rubric Avg rose by 6.25 points, Task Avg by 6.32 points, and WebArena-Verified Overall success by 13.33 points.

Why it matters

The broader contribution is a change in the economics of environment creation. If website replicas can be generated, checked, and repaired through an automated pipeline, researchers can assemble a wider range of training and evaluation tasks without treating every website as a bespoke engineering project. This could make web-agent research less dependent on a small set of benchmark sites.

The approach is not equivalent to reproducing the entire live web. Its coverage depends on browser-observable information, captured interaction outcomes, and the verifier’s ability to identify discrepancies. Authentication, changing external data, permissions, and other dynamic behavior may still affect stability. WebAnything is therefore best viewed as a scalable construction workflow for source-grounded environments, rather than a complete simulation of internet behavior.

Source: Hugging Face Daily Papers