Researchers have introduced PatchScout, a multi-agent framework designed to gather reusable collections of Web documents around a defined theme. Unlike systems built to answer a single query, it aims to find relevant material spread across diverse sources, including parts of the web that are not connected through links.
The research frames collection as a long-horizon information-foraging problem. A search query opens a “semantic patch” of results and nearby pages; the system must decide how long to explore that area before searching for another. PatchScout combines adaptive search with local link traversal and uses specialized agents for discovering patches, assessing relevance, exploiting promising areas and deciding when to switch.
The paper argues that existing crawling methods can miss relevant sources when they are disconnected from initial pages, while search-driven agents may stop once they have enough information to answer a question. PatchScout instead treats both the depth of collection within discovered sources and coverage across sources as goals under a fixed interaction budget.
In experiments on four live-web topics, the authors report that PatchScout improved document yield and domain coverage over traversal-based and search-augmented reasoning baselines. In an applied social science task involving an incomplete collection, it found 102 entities that had not previously been identified.
Sources: arXiv cs.AI.