WebAnything Clones 244 Websites on Its Own to Train Web Agents

To teach an AI agent to use websites, you need a playground where it can click, make mistakes and start from scratch. The live internet is a poor fit: its state can't be reset, results can't be reproduced, and one extra click could place a real order. Researchers from Nanyang Technological University and Shanghai Qiji Zhifeng have introduced WebAnything. The system takes an ordinary site address and builds a working local copy of the site by itself, with no manual tuning for each site. The collection currently holds 244 such copies across 14 domains and 53 categories. They contain 12 984 verified tasks for agents.

What's wrong with existing approaches

The authors describe three approaches, each with its own weakness.

• Copies built by people. They are accurate and controllable, but each site needs separate engineering work. One example is WebHarbor, where humans build copies together with a coding agent.
• Generating a site from a description. It scales well, but the result can drift away from the real site.
• Record and replay. The look of the site is preserved, but the agent can't go beyond what was recorded.

WebAnything tries to take the best from each: automation, fidelity to the original and freedom of action.

One agent builds, another checks

The system can't see a site's server-side code, only what the browser shows. The first AI agent, the Explorer–Builder pair, walks through the pages and gathers evidence: page structure, styles, element geometry, screenshots, fonts and images. The key part of the collection is a before-and-after record of every action. A closed and an open menu at the same address count as different states. From this data, the agent writes the code: pages, routes and transitions between states.

The second AI agent, an independent verifier, then takes over. It receives the frozen version and, in a separate session, re-runs the pages, elements, scenarios and data persistence. It compares the results against the records of the original. It is not allowed to edit the code; only the builder can do that.

A discrepancy it finds follows one of two paths:

• If there is enough data and the copy got it wrong, the builder receives a precise description: what was done, what was expected, what actually happened. The builder then fixes the code locally.
• If there isn't enough data, the system returns to the original site and collects what is missing.

After any fix, the version is frozen and checked again. An example from the paper: in Figma, the comment field on the canvas ended up in the wrong place and looked different. The builder changed the markup and styles, and the verifier confirmed the position and visibility.

A finished copy ships with fixed initial data, state reset and result checking. It can be run any number of times.

What the measurements showed

For 85 sites from the historical collection, no manual implementation, debugging or per-site fixes were recorded. Only initial setup and authorization are needed. The authors accounted for model-call costs on 84 sites and describe the total as a lower bound.

They compared quality with WebHarbor on five sites and 14 page states. WebAnything won on all three automatic metrics: pixel match, SSIM structural similarity and Text F1 text accuracy. In a blind evaluation by three annotators, it recorded 12 wins, two ties and no losses. Separately, the team checked 18 discrepancies on six sites, and all of them were confirmed and fixed.

Training usefulness was tested on Qwen3.5-9B. The model was fine-tuned on 5 856 trajectories (60 517 step-by-step examples, 473 training steps). The gains on three external benchmarks:

• REAL, a benchmark built on controlled copies of websites: +9.9 pp in successful tasks.
• Odysseys-Easy: +6.25 pp on the rubric score and +6.32 pp on the share of completed tasks.
• WebArena-Verified: +13.33 pp in overall success rate.

The new 6 812 tasks across 229 sites were not used in this fine-tuning.

What this means in practice

For teams building web agents, the main point is that site copies stop being manual work. Previously, a scenario like "a hundred different shops, banks and task trackers" ran into person-hours. Now it is a question of model-call budget.

There are limitations, too. A site is copied within the declared scope. Hidden server-side logic is reconstructed only from observed behavior. The gain was measured on one small model and a fixed agent harness. It is not a guarantee of the same effect on other models and other tasks.