Context
SearchOS-V1 is an experimental multi-agent web information-seeking framework that represents tasks as relational schema completion with cell-level source citations. Its shared search state tracks evidence, coverage gaps, pending tasks and failed searches; benchmark evaluations report improved completeness against several agent baselines.
What changed
The paper introduces Search-Oriented Context Management (task frontier, evidence graph, coverage map and failure memory), middleware-based monitoring, pipeline-parallel search dispatch and reusable search/access skills. On WideSearch, item F1 was 80.3 versus 76.0 for the strongest baseline; on GISA set questions, F1 was 76.5 versus 63.1, under reported Max@3 evaluation.
Why it matters
Long-horizon AI research agents may fail to expose sources or assemble grounded answers when they lose track of missing facts or repeat searches. The work gives a testable framework for measuring discovery completeness, fine-grained evidence provenance and search-state management beyond simple retrieval or citation counts.
Visibility OS interpretation
Visibility OS interpretation: candidate-source retrieval, source reading, evidence coverage and final cited answer completeness are separate stages. Shared state and gap-aware search policies may influence which evidence is encountered, but this paper does not demonstrate ranking or citation gains in commercial AI search systems.
What remains uncertain
Will gains persist across average-run rather than best-of-three performance, other languages, real-time web changes, or deployed consumer AI search engines? What are the cost and quality tradeoffs against newer search agents under equal budgets? Does stateful evidence tracking change commercial citation exposure?