SearchOS-V1: Towards Robust Open-Domain Information...

SearchOS-V1: Towards Robust Open-Domain Information... | AI Research

Key Takeaways

SearchOS-V1 addresses a critical bottleneck in AI-driven web research: the tendency for information-seeking agents to lose track of their progress, repeat fa...
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.
We introduce SearchOS, a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state.
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.

Paper AbstractExpand

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.

SearchOS-V1 addresses a critical bottleneck in AI-driven web research: the tendency for information-seeking agents to lose track of their progress, repeat failed search patterns, and waste resources as tasks become more complex. By treating search as a structured, system-level process rather than a series of disconnected chat interactions, the framework ensures that agents maintain a clear, shared understanding of what information has been found and what remains to be discovered.

Transforming Search into a Structured Task

The core innovation of SearchOS is the formulation of open-domain research as "relational schema completion." Instead of asking an agent to simply "find information," the system breaks the request down into a database-like structure of tables, entities, and attributes. Each piece of information found is anchored to a specific source and excerpt, creating a verifiable citation matrix. This approach turns vague, open-ended research goals into concrete, measurable targets, allowing the system to track exactly which parts of a request are fulfilled and which are still missing.

Managing State and Collaboration

To prevent agents from working in silos or duplicating efforts, SearchOS introduces Search-Oriented Context Management (SOCM). This system acts as a persistent "memory" shared across all agents, consisting of four key components:

Frontier Task: A pool of pending work that tracks dependencies and prevents redundant searches.
Evidence Graph: A repository of atomic findings that preserves the provenance of every piece of data.
Coverage Map: A real-time dashboard that measures progress toward completing the required schema.
Failure Memory: A record of unsuccessful search paths that prevents agents from repeating the same dead ends.
By using a pipeline-parallel scheduling mechanism, the system keeps agents busy by continuously assigning them new, unresolved gaps as soon as they finish a task, significantly improving overall efficiency.

Middleware for Reliable Execution

SearchOS includes a "Search Tool Middleware Harness" that acts as a guardrail for agent behavior. Because individual AI models can sometimes struggle to follow instructions as a conversation grows long, this middleware sits outside the agent's prompt. It intercepts interactions to automatically extract evidence, enforce budget limits, and detect when a search has stalled. By handling these administrative and safety tasks at the system level, the agents are free to focus entirely on the logic of the search itself.

Proven Performance

The framework also utilizes a hierarchical skill system, which separates high-level search strategies from specific, technical procedures for accessing different types of websites. In evaluations on the WideSearch and GISA benchmarks, SearchOS outperformed existing single- and multi-agent systems across all reported metrics. By providing a robust, observable, and stateful environment, SearchOS demonstrates a more reliable path for AI agents to perform complex, long-horizon research tasks.