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

TL;DR - SearchOS-V1 is a multi-agent framework for open-domain information seeking that makes search progress explicit, persistent, and shared to prevent agents from getting stuck in repetitive, budget-wasting loops. It matters because it targets a core failure mode of tool-integrated LLM search agents: losing track of task progress as interaction histories grow.

• Reframes open-domain information seeking as relational schema completion with grounded citations—agents discover entities, fill attributes across linked tables, and anchor each value to source evidence.
• Introduces Search-Oriented Context Management (SOCM) that externalizes evolving state into four structures: a Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory.
• Uses pipeline-parallel scheduling to overlap sub-agent execution and continuously refill freed slots with tasks targeting unresolved coverage gaps, improving utilization/throughput.
• Adds a Search Tool Middleware Harness that intercepts model/tool interactions to record evidence, react to stalls or budget exhaustion, and reuse hierarchical strategy/access skills; reportedly leads all metrics over single- and multi-agent baselines on WideSearch and GISA (specific numbers not provided).