WebAnything: From the Wild Web to Scalable Agent Environments

WebAnything automatically reconstructs source-grounded executable web environments from real URLs using an Explorer–Builder and independent verifier, yielding 244 artifacts and 12,984 accepted tasks and improving agent benchmarks after SFT.

Web agents need executable, resettable website environments for reproducible training and evaluation, but human-assisted mirrors require per-site engineering and generative construction can diverge from specified real websites. Capture-and-replay preserves observations, yet interaction coverage is bounded by captured content and replay capabilities. The paper identifies a Source-Grounding Gap, where inaccessible backends and website internals make structures, resources, and behaviors hard to reconstruct, and a Verification and Feedback Gap, where builder self-checks can miss source-inconsistent defects and discrepancies may indicate either faulty implementation or insufficient observation.

WebAnything takes a real URL, permitted access scope, and construction budget and seeks a local environment preserving observed presentation, interactions, and state changes under a site contract. An Explorer–Builder collects browser-observable page structure and appearance, resources and dependencies, and before/after action observations, then implements executable local routes and state transitions. An independent verifier checks frozen candidates in a separate context against source evidence and site contracts; evidence-supported implementation errors trigger local repair, while unresolved evidence gaps trigger targeted source recapture. Updated candidates are frozen and reverified before release. Released environments package local implementations, resources, initial data, runtime configuration, and reset interfaces, with shared state and task evaluators for repeatable interaction and trajectory collection.

• WebAnything constructs a resource pool of 244 website artifacts across 14 application domains and 53 categories and supports 12,984 executed and accepted tasks. §Abstract
• On five matched WebHarbor sites and 14 historical states, WebAnything achieves higher aggregate scores on three fidelity metrics and records 12 wins, two ties, and no losses in a blind audit. §Abstract
• Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improves REAL success from 24.2 to 34.1, Odysseys-Easy Rubric Avg from 19.71 to 25.96, Task Avg from 18.86 to 25.18, and WebArena-Verified Overall from 16.41 to 29.74. §4.4
• In the five-site VeriEnv comparison, selected WebAnything pool artifacts achieve 22/22 capability coverage, 29/31 executed-test success, and 29/33 all-test pass, versus VeriEnv 20/22, 21/28, and 21/33. §A.7
• Among 219 builder-pass pairs, independent verification records 56 passed, 139 failed, and 24 source blocked, with 39 evidence-supported discrepancies spanning 22 sites. §B.7
• For the historical 85-site subset, reports record 59 passed sites, 15 source-blocked sites, 6 failed sites, and 5 sites with missing reports; recovered quantifiable model costs total $220.62, with mean per-site cost about $2.63 and median about $1.60. §A.1

• No unified construction success rate is computed because neither the 244-site pool nor the historical 85-site subset has a complete attempt cohort under uniform acceptance criteria (§A.1).
• Recorded model costs are only a lower bound and exclude unpriced calls, browser execution, and other system overhead; they are not extrapolated to 244 sites and are not interpreted as a controlled Verifier ablation (§A.1).
• In the five-site WebHarbor subset, WebHarbor achieves higher capability coverage and interaction-test completion rates than WebAnything (§A.7).
• Interaction recovery is adapter-level and does not imply native reset interfaces, with two additional planned VeriEnv persistence checks unexecuted; Text F1 also uses replay-derived source text with JavaScript disabled versus JavaScript-enabled mirror captures, which can affect scores (§A.7, §A.3).

• The source evidence is defined as page structure and appearance, resources and dependencies, and before/after action observations. §3.1
• Explorer–Builder automatically reconstructs executable environments from real URLs without per-site manual implementation or debugging in audited runs. §1
• Independent verification checks frozen candidates in a separate context and directs discrepancies to local repair or targeted recapture. §1
• Released environments include local implementations, resources, initial data, runtime configuration, reset interfaces, shared state, and task evaluators. §3.4
• Tasks execute through a fixed browser-based computer-use agent harness, and evaluators check final answers or terminal environment states. §3.4
• Accepted trajectories are converted into step-level training examples, with task generation, acceptance, collection, and export counted separately. §3.4
• The historical 85-site subset has no recorded per-site manual implementation, debugging, or repair after initial setup and authorization. §A.1
• The blind audit includes all five sites and 14 page states with randomized A/B order and three annotators, without sampling based on automatic scores. §D