Web agents stop juggling three access modes: one interface unifies search, clicking and revisiting, topping six benchmarks at lower cost

Web access for agents collapses from three parallel schemes into one interface: search, browser actions and filesystem-style selective reading no longer run separately but are carried out by the same objects and operations within a single task trajectory. Search-augmented agents worked through search APIs and cleaned text, browser-based agents through page observations and actions, and filesystem-based agents stored webpages in a workspace and inspected them selectively with shell operations such as grep and read, each covering only information gathering or page interaction. Across six benchmarks, WideSearch, SealQA, WebArena-Verified, WebWalkerQA, DeepResearchBench and DeepWeb-Bench, the interface scores above HybridSBF, Fetch-then-Explore and CLI-Anything Browser, for example 0.782 item F1 on WideSearch and 31.11 overall on DeepWeb-Bench.

Webpages are compiled into addressable information objects and action objects, and the agent uses addresses to specify what to inspect or which action to execute while the interface performs the operation and returns the outcome. WebContract defines the common schema for objects and pages, WebCompiler compiles fetched webpages into operable representations using deterministic rules derived from HTML semantics, and WebCache retains the compiled pages for selective access throughout a task. Each object combines an address, a value and associations that preserve the context and references needed to interpret the value; compilation relies on no model inference and takes 0.46 to 2.79 seconds per page on average across the benchmarks.

The interface targets web agents that must gather information, operate pages and revisit evidence within a single trajectory, and the experiments cover six benchmarks across web search, web navigation and deep research, all with gpt-5.6-luna as the main agent and turn budgets of 100, 50 and 200 for web search, web navigation and deep research. Compilation overhead is measured separately as the time to compile acquired content and store the resulting objects, averaging 0.46 to 2.79 seconds per page across benchmarks.

The analysis notes that when an evidence search does not return the required information, agents may be uncertain about where to look next, and the object association field may guide further inspection, a role left for later examination. Unaccessed content in the cache can still contain useful evidence: when the prefetched collection grows, 9.5% of the gold information in WebCache has not reached the agent's observations, which motivates cache management that preserves potentially useful evidence. Action execution also depends on the browser state currently available, and if the requested control cannot be confirmed, the interface returns the current page information without executing the action.