Why this matters
As information-seeking agents scale their interactions with web tools, latent progress tracking and repeating failed searches become the dominant bottlenecks: agents waste budget on loops and fail to assemble complete, grounded answers. SearchOS reframes open-domain search as relational schema completion with grounded citations and shifts progress from ephemeral chat history into a structured, reusable state that agents can jointly read and update.
Key Findings
- Externalized state: Search-Oriented Context Management (SOCM) stores ongoing work in four persistent artifacts — Frontier Task (next search intents), Evidence Graph (sourced facts and links), Coverage Map (which schema cells are filled) and Failure Memory (recorded dead-ends). This converts transient signals into actionable, shared memory so different sub-agents don’t re-explore the same failures.
- Scheduling and throughput: a pipeline-parallel scheduler overlaps sub-agent execution and refills freed slots with tasks targeting unresolved coverage gaps, increasing utilization and reducing idle tool/query slots compared to naive multi-agent runs.
- Middleware and skills: a Search Tool Middleware Harness intercepts model-tool interactions to ground evidence and detect stalls or budget exhaustion; a hierarchical skill system (strategy + access skills) guides agents to avoid repeating failed patterns across runs.
- Empirical gains: on WideSearch and GISA benchmarks, SearchOS outperforms evaluated single- and multi-agent baselines across coverage and grounded-evidence metrics, showing the architecture improves both completeness and evidence quality.
Who it's for & trade-offs
Great fit if you build or evaluate LLM-based information-seeking agents that must assemble grounded answers from heterogeneous web evidence and need durable cross-agent state to avoid wasted queries. The design is particularly useful for systems where search budgets and tool-call latency matter, and where producing source-linked relational outputs is a requirement.
Look elsewhere if your use case is a single-turn retrieval task, extremely low-latency queries where any external state persistence is unacceptable, or you cannot afford the extra engineering to manage persistent evidence graphs and middleware: SearchOS adds orchestration complexity in exchange for coverage and reuse.
Where it fits
SearchOS sits above individual LLM agents and search tools as a coordination and context layer: think orchestration + persistent state for multi-agent retrieval-augmented workflows. It complements RAG-style pipelines by making coverage and failure history first-class, rather than implicit in prompt/context.
Core mechanism (brief)
The paper formalizes open-domain seeking as filling relational tables with values linked to sources, and implements SOCM plus a pipeline scheduler and middleware harness to record grounded evidence and dynamically reassign tasks. The combination reduces redundant searches and improves the chance of closing coverage gaps across long, multi-step information tasks.
