Large language model (LLM) agents are increasingly acting as direct users of the open web. As tasks become more complex and information‑intensive, reliable web access emerges as a core capability for agentic problem solving. Existing solutions fall into three access models: search‑augmented agents, browser‑based agents, and the newer filesystem‑based agents. The former excel at information retrieval, the latter at interactive page actions, while filesystem agents map webpages to local files. In real‑world long‑horizon, end‑to‑end tasks, users often need to gather broad information, perform browser actions, and revisit collected evidence within a single trajectory. Our preliminary analysis shows that current models either support only information gathering or only interaction; combining both incurs significant cost or performance trade‑offs. To address this, we introduce **Web$hell**, an interface that unifies web access under a single model. A common schema defined by `WebContract` allows `WebCompiler` to compile live webpages into addressable information and action objects, which are stored in `WebCache` and manipulated via a shared set of operations, providing agents with consistent, operable, and persistent web access. Extensive experiments on web search, navigation, and deep‑research benchmarks demonstrate that Web$hell achieves strong task performance at lower cost. Our findings further highlight the importance of interface design for effective web access, and Web$hell opens new possibilities for more capable and efficient LLM agents on the web.
Review