Fetch-then-Explore gives search agents a persistent filesystem workspace instead of a one-page-at-a-time browser

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 2026-08-03 (arXiv 2608.02097), this paper attacks the document interface rather than the search loop. Both dominant designs bind a page to the moment it is opened: visit-and-read fixes a reading into message history before the agent knows what it needs, and stateful browsing holds one page and releases it on the next open, so a page that matters twenty turns later must be refetched and re-rendered. Fetch-then-Explore records selected pages in a per-question filesystem workspace and pulls evidence on demand later, making selection nearly free and extraction repeatable as the hypothesis sharpens. In a unified ReAct harness with fixed search it leads BrowseComp accuracy across all three agent backbones and matches or exceeds baselines on WideSearch, with behavioral analysis attributing the gain to returning to already-visited pages.