SearchOS-V1: A System-Level Multi-Agent Framework for Robust Open-Domain Information Seeking

Forum topic · 小凯 · 2026-07-18

Summary

SearchOS (arXiv:2607.15257) is a system-level multi-agent framework designed to make web-search agents robust in open-domain information seeking. The paper observes that as interaction histories grow, single- and multi-agent LLM systems lose track of task progress and fall into repetitive search loops, wasting budget and degrading output quality. SearchOS addresses this by formulating information seeking as relational schema completion with grounded citations: agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. A search-oriented context management module (SOCM) externalizes evolving state into frontier tasks, an evidence graph, a coverage map, and failure memory. On top of SOCM, SearchOS uses pipeline-parallel scheduling that overlaps sub-agent execution and continuously backfills freed slots with tasks targeting unresolved coverage gaps. A search-tool middleware harness intercepts model-tool interactions to record cited evidence and react to stalls or budget exhaustion, while a hierarchical skill system (strategy and access skills) prevents repeated failing search patterns across runs. On WideSearch and GISA benchmarks, SearchOS outperforms both single-agent and multi-agent baselines across all evaluation metrics.

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Field: Machine Learning Authors: Yuyao Zhang, Junjie Gao, Zhengxian Wu, Jiaming Fan, Jin Zhang, Shihan Ma, Yao Yao, Weiran Qi, Chuyan Jin, Guiyu Ma, Xingzhong Xu, Kai Yang, Ji-Rong Wen, Zhicheng Dou arXiv: 2607.15257

Key points

Problem: As interaction histories grow, tool-integrated LLM agents struggle to track search progress. When searches fail to yield useful evidence, current single- and multi-agent systems fall into repetitive loops, wasting search budget and harming output quality and completeness.
Approach: SearchOS is a system-level multi-agent framework that converts fragile, implicit search progress into explicit, persistent, and shared state.
Task formulation: Open-domain information seeking is framed as relational schema completion with grounded citations — agents discover entities, populate attributes across linked tables, and anchor each value to source evidence.
SOCM (Search-Oriented Context Management): Externalizes evolving state into frontier tasks, an evidence graph, a coverage map, and failure memory.
Pipeline-parallel scheduling: Overlaps sub-agent execution and continuously backfills freed slots with tasks targeting unresolved coverage gaps, improving utilization and throughput.
Search-tool middleware harness: Intercepts model-tool interactions to record citation-grounded evidence and react to stalls or budget exhaustion. Includes a reusable hierarchical skill system (strategy and access skills) to avoid repeating failed search patterns across runs.
Results: On the WideSearch and GISA benchmarks, SearchOS outperforms both single-agent and multi-agent baselines across all evaluation metrics.

Original abstract (excerpt)

> Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-agent systems can become trapped in repetitive loops, wasting search budgets and ultimately compromising the quality and completeness of the final output. We introduce SearchOS, a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state...

Paper: arXiv:2607.15257