Skip to main content
Back to timeline
arXivSource publication:

One interface instead of three: compiled web objects let agents top six benchmarks at lower cost

Synopsis

AI agents working on the web have had to switch among search interfaces, browser control and filesystem access, each strong at gathering information or at page interaction but costly or weaker when combined. Web$hell compiles live pages into addressable information and action objects handled through shared operations, and reaches the highest task performance on six benchmarks spanning web search, navigation and deep research while using fewer turns and lower inference cost.

Source-provided article image: Web$hell: A Unified Interface for Agentic Web Access
Figure 1 · arXiv

Interpretation

Web access moves into a single interface: instead of calling search, browser and filesystem separately, an agent sends requests against one set of addressable page objects. Search-augmented agents relied on search APIs and cleaned text, browser agents on page observations and actions, and filesystem agents on stored pages inspected with shell operations, so each covered only one side of information gathering or page interaction. Across WideSearch, SealQA, WebArena-Verified, WebWalkerQA, DeepResearchBench and DeepWeb-Bench, Web$hell has the highest task performance, for example a 0.679 success rate on WebArena-Verified versus 0.578 for HybridSBF and 0.264 for Fetch-then-Explore.

The unified interface keeps page content and executable actions in one representation, so an agent that reads information can point directly at the matching action. WebContract defines each object by an address, a value and associations that carry the context needed to interpret the value plus references to related objects. WebCompiler applies deterministic rules derived from HTML semantics, without model inference, to compile captured pages into text, table, link and action objects; WebCache stores the compiled pages and tracks state changes while earlier states stay read-only, at an average compilation cost of 0.46 to 2.79 seconds per page.

The cost advantage is clearest on deep research, where agents gather and synthesize evidence across sources in fewer turns. CLI-Anything Browser used 52.4 turns on deep research and HybridSBF used 30.9, against 17.6 for Web$hell. Web$hell scores 31.11 on DeepWeb-Bench versus 28.37 for HybridSBF and 28.27 for Fetch-then-Explore, and reaches 0.782 item-level F1 on WideSearch versus 0.745 for the strongest baseline, Fetch-then-Explore.

Perspective

The report covers web search, web navigation and deep research tasks, with gpt-5.6-luna as the main-agent model, search through the Serper API, and pages containing reference answers excluded to prevent leakage. Web navigation is evaluated mainly on self-hosted WebArena-Verified sites and on WebWalkerQA starting URLs, a controlled evaluation setting. Compilation overhead averages 0.46 to 2.79 seconds per page, while page loading time varies with site rendering and is not used as a direct measure of interface overhead.

As the number of prefetched pages grows, the fraction of cached pages the agent inspects falls from 14.7% to 10.4%, and 9.5% of the gold information in the cache never reaches the agent's observations, leaving cache management that preserves potentially useful evidence an open question. Action execution depends on the browser state currently available, and when the tab for a saved interaction state is no longer available the interface reports that the action cannot be executed directly from that state.