Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - An arXiv preprint proposing "Fetch-then-Explore," a document interface for search agents that decouples page selection from evidence extraction by persisting fetched pages in a per-question filesystem workspace. It matters because it addresses an under-studied bottleneck — how agents read pages, not how they find them — and improves long-horizon web research accuracy.
- Critiques the two dominant interfaces: visit-and-read (dumps a fixed page reading into context at fetch time, before the agent knows what it needs) and stateful browsing (extracts on demand but holds only one page and drops it when the agent navigates away); both force re-fetching pages that matter later.
- Fetch-then-Explore stores selected pages on the filesystem rather than in the context window or a transient session, making selection nearly free and letting extraction be deferred and repeated as the agent's hypothesis sharpens.
- Evaluated in a unified ReAct harness with fixed search against snippet-only, visit-and-read, and browsing baselines on BrowseComp and WideSearch across three agent backbones: leads BrowseComp accuracy at every backbone and generally matches or exceeds baselines on WideSearch.
- Behavioral analysis attributes the gains to page revisitation — the workspace returns to previously opened pages far more often than transient interfaces, allowing evidence missed on a first pass to be recovered.
Sources (1)
Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents
TL;DR - An arXiv preprint proposing "Fetch-then-Explore," a document interface for search agents that decouples page selection from evidence extraction by persisting fetched pages in a per-question filesystem workspace. It matters because it addresses an under-studied bottleneck — how agents read pages, not how they find them — and improves long-horizon web research accuracy.
- Critiques the two dominant interfaces: visit-and-read (dumps a fixed page reading into context at fetch time, before the agent knows what it needs) and stateful browsing (extracts on demand but holds only one page and drops it when the agent navigates away); both force re-fetching pages that matter later.
- Fetch-then-Explore stores selected pages on the filesystem rather than in the context window or a transient session, making selection nearly free and letting extraction be deferred and repeated as the agent's hypothesis sharpens.
- Evaluated in a unified ReAct harness with fixed search against snippet-only, visit-and-read, and browsing baselines on BrowseComp and WideSearch across three agent backbones: leads BrowseComp accuracy at every backbone and generally matches or exceeds baselines on WideSearch.
- Behavioral analysis attributes the gains to page revisitation — the workspace returns to previously opened pages far more often than transient interfaces, allowing evidence missed on a first pass to be recovered.