Search Agents Need an 'Operating System': SearchOS Open-Sources Multi-Agent Collaboration Framework

Over the past year, the capability improvement of search agents has mainly followed a clear path: through reinforcement learning, models learn when to search, how to construct queries, how to call tools, and how to reason based on page content. However, improvements in model search capabilities do not naturally solve system-level problems in long-horizon tasks.

In complex information retrieval processes, models often need to dynamically discover hundreds of entities, fill in numerous attributes, and verify facts across multiple sources. Models with limited context windows often struggle to complete such tasks. If plans, progress, evidence, and failure records are mainly stored in the ever-growing conversation context, several issues arise as tasks lengthen: confirmed facts are lost due to compression, different subtasks repeat queries, multiple agents have inconsistent understandings of field definitions, sources become decoupled from conclusions, and the system struggles to accurately determine 'how much is done and what is missing.'

The latest research work from Renmin University and Ant Group, SearchOS, provides an answer by making search state a schedulable infrastructure and offering system-level multi-agent search collaboration.

SearchOS first draws on the design philosophy of relational databases to redefine the open-domain information retrieval problem. Given a natural language request q, the system first creates a relational search schema, where T_m defines the table schema, A_m are the columns to fill, P_m are the primary keys distinguishing each row, and R are the correspondences between tables. Simple tasks may have only a few rows; complex tasks can have multiple tables connected via primary and foreign keys.

Around this definition, SearchOS designs Search-Oriented Context Management (SOCM), maintaining a retrieval task as four types of continuously evolving shared state: Frontier Tasks (saves pending tasks with priorities and dependencies), Evidence Graph (records findings, sources, confidence, and relationships between evidence), Coverage Map (real-time completion status of the relational schema), and Failure Memory (records invalid queries, inaccessible sources, and missing capabilities). This mechanism can be understood as a shared progress table for all agents.

At runtime, a long-lived Orchestrator handles unified planning, scheduling, and convergence. It first goes through an Explore phase to understand the problem and discover candidate entities, then establishes a search schema including primary keys and inter-table relationships, decomposes entities, attributes, or relationships not yet filled into tasks, and dispatches them to multiple short-lived Search Agents. Scheduling uses pipeline parallelism and continuous dispatch, treating each search subtask as a micro-batch, with different tasks staggered and overlapping to improve agent slot utilization and overall throughput.

To ensure control logic does not rely on agents remembering and executing on their own, SearchOS adds a Search Tool Middleware Harness between models and tools, including Context Middleware (injects relevant state on demand and controls context size), Sensor Middleware (identifies loops, repeated queries, and stalls), and Evidence Extraction Middleware (structured extraction, unit normalization, citation anchoring, and evidence storage). Each Search Agent only needs to handle local search subtasks; state management, evidence processing, and anomaly intervention are handled uniformly by the system layer.

SearchOS also builds a hierarchical skill system for search agents, categorizing skills into orchestration, strategy, and access types. In the first open-source version, the authors pre-built approximately 280 skills. Strategy skills encapsulate methods for 'how to search' such as ranked retrieval, multi-hop search, and entity disambiguation; access skills handle 'how to access' issues like anti-crawling, login walls, dynamic pages, and deep directories. Skills can be routed and reused based on tasks, and can evolve from successful and failed search trajectories.

The authors also point out that the core contribution of SearchOS-V1 is how to abstract intermediate products during the search process into system state shared across agents and provide infrastructure for task execution. How to automatically generate large-scale skills from data sources, interaction trajectories, and user intentions is explicitly left for future work to introduce and discuss.

To evaluate the effectiveness of SearchOS, the authors conducted evaluations on two open information retrieval benchmarks. The first is WideSearch, targeting large-scale wide-table information collection, containing 200 manually crafted questions (100 in Chinese, 100 in English) across 15+ domains, with answers required to be verifiable complete tables. The second is GISA, with 373 questions simulating real retrieval scenarios, covering answer formats including single items, sets, lists, and tables, focusing on comprehensiveness of open-world information collection.

According to the project's published max@3 results, SearchOS leads all evaluated single-agent and multi-agent baselines across all F1 metrics. Specifically, WideSearch Item F1 is 80.3, Row F1 is 56.5; GISA Table Item F1 is 76.9, Set F1 is 76.5. On Set F1, which requires enumerating complete sets, SearchOS is 13.4 points higher than the next best baseline, with gains mainly from Coverage Map-driven continuous gap filling.

To further verify the flexibility of the relational search schema, the authors also conducted experiments on 40 questions that could be decomposed into multiple tables. They first used GPT5.5 to pre-build single-table and multi-table structures and performed collection tasks. Results showed that even with manually pre-specified optimal table structures (Oracle) for each question, Item F1 was still 8.2 points lower than SearchOS, and Row F1 was 7.7 points lower. This indicates that in real scenarios, different information collection tasks require different table structures, thus validating the effectiveness of SearchOS's dynamic creation of table structures based on discovered entity relationships during exploration.