SearchOS: Shared State for Multi-Agent Web Search

SearchOS reframes open-domain web search as citation-grounded schema completion, using shared state, middleware, scheduling, and reusable skills to coordinate search agents.

TL;DR — SearchOS is a system-level multi-agent framework for open-domain web search. It makes search progress explicit through SOCM, which externalizes state into Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory; formulates information seeking as cited relational schema completion; and uses middleware, pipeline-parallel scheduling, and reusable skills to reduce repeated failed search patterns. The abstract claims SearchOS leads all metrics on WideSearch and GISA but gives no exact scores. Some component roles are inferred because the abstract names the components without fully defining their internals.

Problem: search agents lose durable state

SearchOS addresses a failure mode in tool-using language-model agents: as a search session grows, agents can lose track of what they have tried and what remains unresolved. The abstract states that "as interaction histories grow, agents increasingly struggle to track task progress."

The core failure is repetition after unsuccessful search. The abstract says that when search attempts fail to yield useful evidence, "current single- and multi-agent systems can become trapped in repetitive loops," which wastes search budgets and can compromise output quality and completeness.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.