SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration
arXiv · HuggingFace · ▲60
Abstract (verbatim)
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. First, we formulate open-domain information seeking as relational schema completion with grounded citations, where agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. Then we design Search-Oriented Context Management (SOCM), which externalizes the evolving state into Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory. Built on SOCM, SearchOS applies a pipeline-parallel scheduling mechanism that overlaps the execution of sub-agents and continuously refills freed slots with tasks targeting unresolved coverage gaps to improve utilization and throughput. To schedule and control the execution of search agents, SearchOS introduces a Search Tool Middleware Harness that intercepts model and tool interactions to record grounded evidence and react to stalls or budget exhaustion, and provides a reusable hierarchical skill system comprising strategy and access skills to augment the agents' search process and avoid repeating failed search patterns across runs. On WideSearch and GISA, SearchOS leads all metrics among the evaluated single- and multi-agent baselines, paving the way toward robust information-seeking collaboration.
Background
Background Analysis
Technical Context: With the advancement of tool-integrated large language models (LLMs), information-seeking agents have become capable of web search, browsing, and reasoning to extend knowledge beyond model parameters. Such technologies are widely applied in scenarios requiring multi-source information integration, such as academic research, business analysis, or intelligent assistants. The core need is to enable agents to reliably track task progress, avoid redundant efforts, and generate verifiable answers.
Previous Issues: Existing methods face two major challenges in complex tasks. First, as interaction history grows, agents struggle to track completed and pending tasks, leading to lost evidence, redundant collection, or contradictory conclusions. For example, a single agent may get stuck in repetitive loops when searches fail, wasting resources. Second, simply adding more agents does not solve the problem—parallel workers may duplicate efforts, disagree on objectives, or leave resources idle. The root cause is that traditional approaches treat plans, progress, and failures as transient dialogue content rather than system-level states, making long-horizon tasks unsustainable and untraceable.
Proposed Solution: The paper introduces SearchOS, a framework addressing these issues through:
1. Relational Search Modeling: Framing open-domain information retrieval as "relational schema completion," where entities, attributes, and evidence are explicitly linked, making progress measurable.
2. Explicit State Management: Introducing Search-Oriented Context Management (SOCM) to externalize execution states (e.g., pending tasks, evidence graphs, coverage maps) for shared access across agents.
3. Pipeline-Parallel Scheduling: Adopting a GPU-like pipeline mechanism to dynamically assign unresolved tasks, improving resource utilization.
4. Middleware Control: Using a "Search Tool Middleware Harness" to monitor and regulate agent behavior, preventing repeated errors and enforcing resource limits.
5. Hierarchical Skill System: Separating reusable search strategies from site-specific access skills for cross-task reuse.
Key Differences: Compared to prior work, SearchOS stands out by:
- Shifting from Implicit to Explicit States: Extracting task progress from dialogue history into system-level states, reducing dependency on agent memory.
- Moving from Single-Agent to Collaborative Multi-Agent Systems: Enabling efficient coordination via shared states and dynamic scheduling, rather than naive parallelism.
- Designing Reusable Skills: Creating a hierarchical system for cross-scenario reuse, instead of relearning strategies for each task.
These innovations make SearchOS superior to existing baselines in experiments, laying the groundwork for robust information-seeking collaboration.





