🤖 AI

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

This paper introduces SearchOS, a robust multi-agent framework that enhances open-domain information seeking by externalizing task progress into a shared state via Search-Oriented Context Management, employing pipeline-parallel scheduling to maximize efficiency, and utilizing a middleware harness with hierarchical skills to prevent repetitive search loops and improve evidence coverage.

Original authors: Yuyao Zhang, Junjie Gao, Zhengxian Wu, Jiaming Fan, Jin Zhang, Shihan Ma, Yao Yao, Weiran Qi, Chuyan Jin, Guiyu Ma, Xingzhong Xu, Kai Yang, Ji-Rong Wen, Zhicheng Dou

Published 2026-07-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuyao Zhang, Junjie Gao, Zhengxian Wu, Jiaming Fan, Jin Zhang, Shihan Ma, Yao Yao, Weiran Qi, Chuyan Jin, Guiyu Ma, Xingzhong Xu, Kai Yang, Ji-Rong Wen, Zhicheng Dou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a massive mystery, but instead of a single notebook, you have a team of junior detectives. In the world of artificial intelligence, these "detectives" are called agents. They are smart computer programs that can talk to the internet, read websites, and gather information to answer your questions. For a long time, these agents were like detectives who kept forgetting what they had already found. As the mystery got bigger, they would get confused, repeat the same questions over and over, or get stuck in a loop, wasting time and energy. They didn't have a way to keep a shared, organized record of their progress.

To fix this, researchers are building better ways for these agents to work together. They use a concept called Large Language Models (think of them as super-smart brains that can read and write) and give them tools to search the web. The big challenge is long-horizon tasks: these are complex jobs that require gathering hundreds of facts, checking if they match, and organizing them perfectly. If the agents lose track of the "big picture," the final answer becomes messy or incomplete. This paper tackles that exact problem: how do we stop these digital detectives from getting lost in their own work and help them collaborate like a well-oiled machine?

The Problem: The "Forgetful Detective" Loop

Imagine you and your friends are trying to build a giant Lego castle based on a vague instruction manual. You all start grabbing bricks, but no one is keeping a list of what you've already found. One friend keeps looking for a red 2x4 brick even though another friend found it ten minutes ago. Another friend gets stuck trying to open a door that's locked, banging on it repeatedly instead of trying a different door.

This is what happens to current AI search agents. As they try to answer complex questions, they lose track of their progress. They get stuck in repetitive loops, wasting their "search budget" (their time and computing power) on dead ends. They might miss important facts because the evidence got buried in a long chat history, or they might argue with each other because they don't have a shared map of what is known and what is missing. Simply adding more agents doesn't fix this; it just creates more chaos, with everyone stepping on each other's toes.

The Solution: SearchOS, the "Operating System" for Search

The authors introduce SearchOS, a new framework that acts like a central command center or an "operating system" for these search agents. Instead of letting the agents rely on their fading memories, SearchOS forces them to write everything down in a shared, organized notebook that never gets lost.

Here is how it works, using some fun analogies:

1. The Master Blueprint (Relational Schema)
Instead of just asking a vague question like "Tell me about companies," SearchOS turns the task into filling out a giant, structured spreadsheet. Imagine a detective board where you have specific columns for "Company Name," "CEO," "Revenue," and "Founded Year." The agents' job isn't just to "find info"; it's to fill in every single empty cell in this spreadsheet. If a cell is empty, the system knows exactly what is missing. This turns a messy, open-ended search into a clear, measurable game of "fill in the blanks."

2. The Shared Command Center (SOCM)
This is the heart of SearchOS. It's a Search-Oriented Context Management system that acts as a shared brain for the whole team. It keeps four critical things updated in real-time:

The Frontier Task List: A to-do list of exactly which spreadsheet cells are still empty.
The Evidence Graph: A map of every fact found, linked directly to the website where it came from (so you can check the source).
The Coverage Map: A visual dashboard showing which parts of the spreadsheet are filled, which are uncertain, and which are missing.
Failure Memory: A "do not try this again" list. If one agent tries a search and it fails (like a broken website), this memory tells the other agents, "Don't bother trying that again; it's a dead end."

3. The Assembly Line (Pipeline-Parallel Scheduling)
Old systems often made agents wait for each other, like a line of people waiting for a single cashier. SearchOS uses pipeline-parallel scheduling. Imagine a factory assembly line: as soon as one agent finishes a task and frees up a spot, the system immediately hands a new task to the next available agent. No one sits idle. If one agent is slow, the others keep working on different parts of the spreadsheet. This makes the whole team much faster and more efficient.

4. The Referee (Middleware Harness)
Sometimes agents get stuck in a loop or run out of time. SearchOS has a Search Tool Middleware Harness that acts like a strict referee. It watches the agents' every move. If it sees an agent banging on a locked door (a failed search) or repeating the same question, the referee steps in, stops the agent, and gives it a new strategy. It also ensures that every fact the agent finds is "grounded" (linked to a real source) before it gets written into the final report.

5. The Toolbox (Hierarchical Skills)
SearchOS comes with a library of pre-made skills. Think of these as specialized tools in a detective's kit. Some tools are for general searching (like "how to ask a good question"), while others are specific to certain websites (like "how to navigate a government database"). The system teaches the agents how to use these tools so they don't have to reinvent the wheel every time they visit a new site.

What They Found: The Results

The researchers tested SearchOS on two difficult benchmarks called WideSearch and GISA. These tests involve complex questions that require gathering many facts from different sources.

The results were impressive. SearchOS beat all the other methods they tested, including single-agent systems and other multi-agent teams.

On the WideSearch test, it achieved an Item-level F1 score of 80.3, which was the highest among all methods.
On the GISA test, it scored 76.5 on the Set F1 metric, beating the next best system by a huge margin of 13.4 points.

The paper suggests that by forcing agents to work from a shared, explicit state (the spreadsheet and the maps) rather than relying on their internal memories, the system becomes much more reliable. It finds more facts (better recall) without making more mistakes (maintaining high precision).

Why It Matters

This isn't just about making AI faster; it's about making AI trustworthy. When an AI is doing deep research, you need to know that it hasn't missed a crucial fact or made up a number. SearchOS provides a way to see exactly what the AI knows, what it doesn't know, and where it might be stuck. It turns the chaotic process of searching the internet into a structured, observable, and collaborative effort.

The authors suggest that this approach—treating search as a stateful, collaborative process with a shared "operating system"—is a major step forward for building agents that can handle complex, real-world research tasks without getting lost in the noise. They didn't just build a better search engine; they built a better way for AI agents to work together as a team.

Technical Summary: SearchOS-V1

Problem Statement

Recent advances in Tool-Integrated Large Language Models (LLMs) have enabled information-seeking agents to perform iterative web search and reasoning. However, as interaction histories expand during long-horizon tasks, agents struggle to track progress, leading to several critical failure modes:

State Loss: Agents lose track of established facts versus unresolved gaps as evidence becomes buried in context.
Repetitive Loops: When search paths yield no useful information, agents often repeat similar queries or explore dead ends, wasting search budgets.
Inefficient Collaboration: In multi-agent systems, parallel workers may duplicate efforts, disagree on field definitions, or leave execution slots idle while waiting for the slowest task.
Fragile State Management: Conventional agents treat plans, progress, and failures as transient conversation content rather than persistent system state.

The core problem is the lack of a system-level mechanism to externalize search state, making progress explicit, persistent, and shared across agents to prevent redundancy and ensure robustness.

Methodology

The authors introduce SearchOS, a system-level multi-agent framework designed to transform fragile, implicit search progress into explicit, shared state. The methodology is built on four pillars:

1. Relational Search Formulation

The system reframes open-domain information seeking as relational schema completion with grounded citations.

Schema: Tasks are modeled as a set of tables connected by primary-foreign key relations.
Objective: The system must discover entity rows, populate attribute values, and maintain a citation matrix linking every value to a specific source URL and anchored excerpt.
Benefit: This converts open-ended requests into concrete, measurable search targets at the entity and attribute level.

2. Search-Oriented Context Management (SOCM)

To externalize execution state, SearchOS maintains a durable shared state $M_t$ comprising four components, updated via locked read-modify-write operations:

Frontier Task: A dependency-aware task pool representing unresolved schema gaps (e.g., specific cells in a table).
Evidence Graph: A graph storing atomic findings (value, source, supporting span) rather than page-level summaries, preserving fine-grained provenance and handling conflicts.
Coverage Map: A materialized view of schema cells tracking status (missing, filled, uncertain, unreachable) to measure progress quantitatively.
Failure Memory: A shared record of unsuccessful actions (e.g., inaccessible sites, repeated queries) to prevent agents from retracing exhausted paths.

Agents receive role-specific projections of this state rather than the full history, ensuring context relevance and avoiding stale information.

3. Pipeline-Parallel Orchestration

Instead of synchronized batch processing, SearchOS employs pipeline-parallel scheduling:

Continuous Dispatch: As soon as an agent slot is freed, it is immediately assigned a new task targeting an unresolved coverage gap.
Utilization: This approach overlaps the execution of sub-agents (Explore, Search, Writer), reducing idle time caused by "stragglers" and increasing overall throughput.
Role Specialization: An orchestrator decomposes the schema and coordinates specialized agents, while a writer synthesizes the final report based on the accumulated evidence.

4. Search Tool Middleware Harness

A system-level harness intercepts model and tool interactions to enforce global invariants without relying on agent prompts:

Context Middleware: Composes role-specific context by injecting relevant state and retrieving skills before model calls.
Evidence Extraction Middleware: Automatically binds and anchors evidence from tool observations to the schema, updating the Evidence Graph and Coverage Map atomically.
Sensor Middleware: Monitors for stalls (lack of coverage growth) and budget exhaustion. It can intervene by injecting corrections, switching strategies, or stopping branches, ensuring robust execution even when agents fail to self-correct.

5. Hierarchical Search Skills

The system utilizes a reusable skill library categorized into:

Orchestrator Skills: Global playbooks for task decomposition and schema alignment.
Strategy Skills: Source-independent methods for query reformulation and reasoning.
Access Skills: Source-specific retrieval procedures (e.g., for government portals or encyclopedias) that handle navigation and parsing.

Key Contributions

Relational Search Formulation: Defining information seeking as relational schema completion with grounded citations, providing a unified objective for entity discovery and evidence attribution.
Stateful, Pipeline-Parallel Collaboration: Introducing SOCM and a continuous orchestration scheme that coordinates agents via shared state and dynamically assigns work around coverage gaps.
Middleware-Governed Execution: Moving evidence processing, context management, and behavioral safeguards (stall detection, budget enforcement) outside of agent prompts via a dedicated middleware harness.
Hierarchical Search Skills: Designing a skill system that separates reusable strategies from site-specific access procedures to augment the search process across sessions.

Experimental Results

SearchOS was evaluated on WideSearch (large-scale fact collection) and GISA (general information-seeking assistant) benchmarks, comparing against single-agent baselines (ReAct, Plan-and-Solve) and multi-agent systems (A-MapReduce, Web2BigTable, Table-as-Search).

WideSearch: SearchOS achieved the best performance across all metrics, reaching 80.3 Item-level F1, outperforming the strongest baseline (A-MapReduce) by +4.3 points. It also led in Row-level F1 (56.5), demonstrating that the recall improvements did not sacrifice precision.
GISA: SearchOS led across all question types, with a particularly significant gain on Set questions (76.5 Set F1), exceeding the best baseline by 13.4 points. This highlights the system's ability to handle complete enumeration tasks effectively.
Ablation Studies:
- Schema Planning: Dynamic schema selection outperformed fixed single-table or multi-table schemas by a wide margin, confirming that task-appropriate structures should be determined during search.
- Pipeline Scheduling: Continuous dispatch reduced end-to-end time by 24.3% and improved slot utilization compared to batch scheduling, while achieving higher F1 scores.
- Middleware Governance: Interventions by the Sensor Middleware successfully redirected stalled search trajectories, restoring progress in early, mid-run, and late stages.
- Skills: Utilizing hierarchical skills improved Item F1 by 2.0 points and Row F1 by 3.4 points, while reducing session time by 36.6% and search calls by 39.1%.

Significance and Claims

The paper claims that SearchOS paves the way toward robust information-seeking collaboration by addressing the fundamental limitations of treating search state as transient conversation history. By externalizing state into a persistent, shared system (SOCM) and governing execution through a middleware harness, the framework ensures that:

Progress is measurable and gaps are explicit.
Redundant efforts and repetitive loops are minimized.
Search budgets are utilized efficiently through pipeline parallelism.
Evidence is rigorously grounded and verifiable.

The authors position SearchOS not merely as a better agent, but as a system-level infrastructure that makes long-horizon search reliable, observable, and scalable, moving beyond the limitations of current single- and multi-agent approaches that rely solely on prompt-based reasoning.