Fetch-then-Explore Gives Search Agents a Filesystem Workspace

Fetch-then-Explore lets AI search agents save pages to a workspace and pull evidence later, rather than reading each page once, per a new paper.

That matters for anyone building a research agent. The design beat both existing approaches, visit-and-read and stateful browsing, on BrowseComp accuracy across all three backbone models tested.

The two current designs both tie a page's contents to the moment the agent opens it. Visit-and-read fixes a page's text into the agent's message history before it knows what it needs from that page. Stateful browsing keeps a page open during a session, but releases it once the agent moves to the next one.

This design skips that tradeoff. It records the pages an agent selects in a workspace tied to the question. Then it lets the agent pull specific evidence from them later, once its read of the question has sharpened. Selecting a page becomes close to free, and extracting from it is repeatable instead of a one-shot read, per the paper.

The take

If a workspace beats two designs across three different backbone models, the design of memory is doing more work than the choice of model. Watch whether BrowseComp gains from here come from bigger backbones or from how agents store what they've already seen. The evidence so far points at storage, not backbone size.