HuggingFace Papers
8/10 signal
SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration
agenticmulti-agentcontextmemorytool-use
Summary
This paper introduces SearchOS, a multi-agent framework designed to make open-domain information-seeking more robust. It addresses the problem of agents getting stuck in repetitive, failed search loops by externalizing search progress into an explicit, shared state. The system uses Search-Oriented Context Management (SOCM) with components like an Evidence Graph and Failure Memory to track progress and a pipeline-parallel scheduler to manage sub-agents. On the WideSearch and GISA benchmarks, SearchOS outperforms all evaluated single- and multi-agent baselines across all metrics.
Problem
As interaction histories with information-seeking agents grow, they 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. This wastes search budgets and ultimately compromises the quality and completeness of the final output.
Method
The authors introduce SearchOS, a system-level multi-agent framework that formulates information seeking as a relational schema completion task with grounded citations. The core of the system is Search-Oriented Context Management (SOCM), which externalizes the search state into a Frontier Task, an Evidence Graph, a Coverage Map, and a Failure Memory. SearchOS uses a pipeline-parallel scheduling mechanism to manage sub-agents and a Search Tool Middleware Harness to intercept tool interactions, record evidence, and deploy a hierarchical skill system to avoid repeating failed searches. The framework was evaluated against single- and multi-agent baselines on the WideSearch and GISA benchmarks.
Details
Benchmark Performance:
- On the WideSearch and GISA benchmarks, SearchOS leads all metrics among the evaluated single- and multi-agent baselines.
Framework Components and Mechanisms:
- Relational Schema Completion: The core task is framed as agents discovering entities, populating attributes across linked tables, and anchoring each value to source evidence.
- Search-Oriented Context Management (SOCM): This system externalizes the evolving state of the search into four explicit, persistent, and shared components: a Frontier Task, an Evidence Graph, a Coverage Map, and a Failure Memory.
- Pipeline-Parallel Scheduling: A scheduling mechanism overlaps the execution of sub-agents and continuously refills freed slots with tasks that target unresolved coverage gaps, improving system utilization and throughput.
- Search Tool Middleware Harness: This component intercepts model and tool interactions to record grounded evidence and react to stalls or budget exhaustion.
- Hierarchical Skills System: The middleware provides a reusable system of strategy and access skills to augment the agents' search process and avoid repeating failed search patterns across runs.
What's new
The paper introduces SearchOS, a system-level multi-agent framework that makes search progress explicit, persistent, and shared. Its primary contribution is the Search-Oriented Context Management (SOCM) system, which externalizes agent state into novel components like an Evidence Graph, Coverage Map, and Failure Memory to prevent repetitive search failures.
Conclusion
The SearchOS framework successfully mitigates the tendency of information-seeking agents to get trapped in repetitive, unproductive search loops. By turning fragile, implicit search progress into an explicit and shared state, the system demonstrates superior performance against existing baselines on the WideSearch and GISA benchmarks. This approach paves the way toward more robust and effective information-seeking collaboration among AI agents.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.