WEBANYTHING · RESEARCH PROJECT

From the wild web
to scalable agent environments.

Automatically reconstructing interactive, source-grounded web environments for agent training and evaluation.

Tianci Xie · Zikun Cai · Chaoyun Zhang · Pinxue Guo · Bingxuan Li · Rui Zheng · Hang Yan

244Website Artifacts
14Application Domains
53Fine-grained Categories
Discover the research

Why WebAnything

A web environment should preserve the task, not just the screenshot.

Static capture misses interaction depth. Hand-built mirrors can scale poorly. WebAnything uses browser-observable evidence to reconstruct the visible state and the behavior required by an agent task.

Figure 1. WebAnything combines broad reconstruction coverage with automated construction and source-grounded fidelity.
01

Visual Similarity

Match the rendered source states that an agent actually observes.

02

Interaction Fidelity

Preserve controls, state transitions, routes, and deterministic reset.

03

Task Preservation

Keep the complete path needed to finish the target workflow.

Method

Explore. Rebuild. Freeze. Verify.

The Builder and Verifier are separated by a frozen candidate boundary. Verification findings can trigger a focused loop without discarding the source evidence already collected.

Figure 2. The WebAnything Builder–Verifier pipeline used in the repository README, including Full/JIT coverage and the implementation-error and evidence-gap feedback loops.

Stage 01

Explore the source through the browser

The Builder follows the target task, records page states, and collects the evidence needed to reproduce the observed path.

  • Interact and navigate
  • Collect page states
  • Record interaction metadata
DOMRendered structureObserved elements and state
UIInteraction metadataControls, routes, and actions
IMGPage snapshotsVisual reference at each step
NETNetwork resourcesSource-observed assets only

Observed task depth

Accepted executions include substantial multi-step interaction.

Statistics below use the 6,172 accepted source tasks whose execution traces were analyzed in the paper.

10.4mean tool steps
8median
2190th percentile
89maximum
Tasks over 10 tool steps34.7%
2,139 of 6,172 accepted tasks

Evaluation

Environment quality and downstream utility.

Results use distinct evaluation settings. Fidelity compares matched historical states; post-training reports the trained model under each benchmark's stated denominator and harness.

Source fidelity · 14 historical states

Higher aggregate fidelity on all three reported metrics.

Pixel similarity0.9240WebHarbor 0.8083
SSIM0.7904WebHarbor 0.6260
Text F10.8541WebHarbor 0.2617
12 wins · 2 ties · 0 losses in the blind overall-fidelity audit
Post-training · Qwen3.5-9B Instruct

Training on WebAnything interactions transfers to external benchmarks.

REAL+9.9pp

22/91 → 31/91, jointly valid tasks

Odysseys-Easy+6.25pp

Rubric Avg, 44 jointly valid tasks

WebArena-Verified+13.33pp

Overall, internal harness

Qualitative comparison. Two historical page states reconstructed by WebAnything and WebHarbor. The full paper reports protocol and scoring details.

Task-driven case study

One URL. One query. A working interaction path.

Exact user query
“在 Mermaid Live Editor 中,把默认流程图改为「草稿 → 审核 → 发布」,确认预览显示三个节点,然后导出 SVG。”

Edit the default diagram to “Draft → Review → Publish,” confirm all three nodes in the preview, then export an SVG.

1 · Open the reconstructed editor 1 / 4
Reconstructed Mermaid editor at task entry
Verifier report · passed

The reconstructed page supports the complete demonstrated task path.

The replay edits the diagram, observes a responsive preview, opens the Actions menu, and downloads an SVG containing the requested labels.

Read the case study
5 / 5task actions passed
0.96964visual comparison
SVGdownload verified

Scope note. The Verifier report passed. The original supervisor run stopped at blocked_retry_limit because its gate rejected structured JSON assertions; the validator and regression test were corrected and the saved Verifier output passed the repaired gate. This case establishes the demonstrated task path, not a fresh end-to-end supervisor pass.

Independent verification

Failures become evidence for the next repair.

The Verifier checks fidelity, functionality, safety, and reproducibility against the frozen candidate. Evidence gaps guide targeted recapture; implementation errors guide local repair.

Verifier cases. Findings from independent verification across layout, control behavior, and navigation defects.

Use WebAnything

Start from a URL and a task.

Install the local skill once, then ask an agent such as Codex to reconstruct a site in Full or JIT mode. The repository includes the pipeline, verifier contracts, demos, and the 244-site metadata inventory.

terminal
# One-time setup
uv sync --locked --extra browser --extra visual --extra builder
uv run python -m playwright install chromium
npm --prefix tools/codex-mirror-trace ci
uv run webanything setup

# Then reconstruct a task path
uv run webanything create https://example.org/ \
  --mode jit \
  --task "Complete the target workflow"

Citation

WebAnything

Read the paper on arXiv:2610.37786 .

@article{xie2026webanything,
  title   = {WebAnything: From the Wild Web to Scalable Agent Environments},
  author  = {Xie, Tianci and Cai, Zikun and Zhang, Chaoyun and Guo, Pinxue and Li, Bingxuan and Zheng, Rui and Yan, Hang},
  journal = {arXiv preprint arXiv:2610.37786},
  year    = {2026}
}
Expanded research figure