Fetch-then-Explore: New Search Agent Method Decouples Page Selection from Evidence Extraction

A recent research article on arXiv (2608.02097) introduces a technique named Fetch-then-Explore designed for search agents. This method distinguishes between selecting pages and extracting evidence, storing pages in a dedicated workspace for each question on the filesystem. The authors contend that existing search agents, which often need to conduct numerous searches to answer questions, depend on document interfaces that link a page to its initial opening. The two current interfaces include 'visit-and-read', which integrates page readings into the message history at the time of fetching, and stateful 'browsing', which retrieves content on demand but can only manage one page at a time. Both methods necessitate re-fetching and re-rendering of pages that may later become pertinent. Fetch-then-Explore mitigates this issue by saving pages in a persistent workspace, enabling agents to access them again without needing to re-fetch. The authors aim to enhance the efficiency and accuracy of multi-step search processes.