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

What Changed

Recent advancements in Tool-Integrated Large Language Models (LLMs) have positioned web search as a fundamental capability for information-seeking agents. However, as the complexity and length of interaction histories grow, these agents frequently encounter difficulties in effectively tracking their task progress. A significant challenge arises when search attempts fail to yield relevant evidence, leading single- and multi-agent systems to become ensnared in repetitive loops. This not only wastes valuable search budgets but also compromises the quality and completeness of the final output.

To address these limitations, researchers have introduced SearchOS, a new system-level multi-agent framework. SearchOS fundamentally alters how search progress is managed, transforming fragile, implicit search states into explicit, persistent, and shared states across agents. This change is critical for mitigating the issues of repetitive searching and improving overall efficiency.

At its core, SearchOS reframes open-domain information seeking as a relational schema completion task, complete with grounded citations. This involves agents identifying entities, populating attributes across linked tables, and meticulously anchoring each value to its source evidence. This structured approach ensures that information is not only found but also properly contextualized and verified.

Technical Details

SearchOS is built upon several key technical innovations. The first is Search-Oriented Context Management (SOCM). SOCM is responsible for externalizing the evolving state of the information-seeking process into four distinct components:

1. Frontier Task: Represents the current set of unresolved tasks or information gaps that need to be addressed.
2. Evidence Graph: A structured representation of discovered entities, their attributes, and the relationships between them, all linked to their source evidence.
3. Coverage Map: Tracks which parts of the information schema have been successfully populated and which still require further investigation.
4. Failure Memory: Records past search attempts that failed to yield useful evidence, preventing agents from repeating unproductive queries or strategies.

This explicit state management allows agents to have a clear, shared understanding of the task's progress, reducing redundancy and improving coordination.

Another critical component is the pipeline-parallel scheduling mechanism. This mechanism enables the overlapping execution of sub-agents. When a sub-agent completes its task or frees up resources, the scheduler continuously refills these slots with new tasks that target unresolved coverage gaps. This dynamic scheduling significantly improves resource utilization and overall throughput, making the information-seeking process more efficient.

To further enhance control and execution, SearchOS incorporates a Search Tool Middleware Harness. This harness intercepts interactions between models and tools, performing several vital functions:

• Recording Grounded Evidence: It systematically records all evidence found during tool interactions, ensuring that all retrieved information is properly attributed and integrated into the Evidence Graph.
• Reacting to Stalls or Budget Exhaustion: The middleware monitors agent activity and resource consumption, allowing the system to detect when agents are stalled or when search budgets are nearing exhaustion. This enables proactive intervention to prevent wasted resources.
• Hierarchical Skill System: The harness provides a reusable hierarchical skill system. This system comprises both strategy skills (high-level approaches to information seeking) and access skills (specific tool usage and query formulation). This hierarchical structure augments the agents' search process, allowing them to learn from past failures and avoid repeating unproductive search patterns across different runs.

Benchmark Analysis

On the WideSearch and GISA benchmarks, SearchOS demonstrated superior performance across all evaluated metrics when compared to existing single- and multi-agent baselines. This indicates a significant improvement in the robustness and efficiency of information-seeking collaboration facilitated by the SearchOS framework.

Developer Implications

For developers working with LLM-powered agents for information retrieval, SearchOS offers a robust framework to overcome common challenges such as repetitive search loops and inefficient resource utilization. The explicit state management via SOCM provides a clear, debuggable view of an agent's progress, making it easier to develop and maintain complex information-seeking applications. The pipeline-parallel scheduling can lead to more efficient use of computational resources and faster task completion, which is crucial for applications requiring rapid information synthesis.

The Search Tool Middleware Harness and its hierarchical skill system provide a structured way to integrate and manage search tools, allowing for more sophisticated and adaptive search strategies. Developers can leverage the Failure Memory to build agents that learn from their mistakes, leading to more intelligent and less wasteful search behaviors over time. The open-domain relational schema completion approach also suggests a pathway for building agents that can not only retrieve information but also structure it into coherent knowledge bases with verifiable citations.

Bottom Line

SearchOS-V1 represents a significant step forward in the development of robust open-domain information-seeking agents. By introducing explicit, persistent, and shared state management through SOCM, coupled with a pipeline-parallel scheduling mechanism and a sophisticated Search Tool Middleware Harness, the framework effectively addresses the limitations of current single- and multi-agent systems. Its ability to formulate information seeking as relational schema completion with grounded citations ensures higher quality and more complete outputs. The demonstrated superior performance on benchmarks like WideSearch and GISA indicates that SearchOS paves the way for more reliable and efficient information-seeking collaboration among AI agents.