SkimSearchAgent

Abstract

Existing deep-research agents typically follow a search-visit workflow that retrieves whole webpages without using the structure already exposed by titles, headings, sections, and metadata. This limits the agent's ability to target specific parts of a page and often sends unnecessary content into the context window. In this work, we introduce Sieve, a search-inspect-fetch strategy built around a Boolean Query Language (BQL). It filters candidates with field-aware Boolean search, ranks them, presents structure-rich inspection cards, and fetches only the selected sections. Across three question-answering collections, Sieve is more accurate than a strong conventional search-visit baseline while using substantially fewer tokens. The implementation is released in the SkimSearchAgent project.