\ul 1]Gaoling School of Artificial Intelligence, Renmin University of China 2]Ant Group \contribution[*]Equal contribution \contribution[†]Corresponding author \contribution[‡]Work done during an internship at Ant Group

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

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

Abstract

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.

\gtechdata

[ Website]https://antins-labs.github.io/SearchOS \gtechdata[ Code]https://github.com/antins-labs/SearchOS \gtechdata[ YouTube]https://www.youtube.com/playlist?list=PLXMUs3Ayz3EQ

1 Introduction

Recent advances in tool-integrated large language models (LLMs) have made web search a core capability of information-seeking agents (Yao et al., 2022; Nakano et al., 2021; Jin et al., 2025a; Li et al., 2026b). These agents can iteratively search, browse, and reason over external sources, extending language models beyond the knowledge available in their parameters. As the field moves toward broad, open-domain and long-horizon tasks (Wong et al., 2025; Zhu et al., 2026), an agent must gather large collections of facts, reconcile evidence across sources, and produce a structured answer with verifiable citations.

Stronger search capability alone, however, does not make this process reliable. As interaction histories grow, agents lose track of what has been established and what remains unresolved. Evidence becomes buried in context, making omissions, redundant collection, and conflicting claims difficult to detect. When a search path yields no useful information, a single agent may repeatedly issue similar searches or continue exploring the same dead end, wasting its budget and degrading the final answer. Simply adding more agents does not eliminate these problems: parallel workers may duplicate effort, disagree on the intended fields, or leave execution slots idle while the system waits for the slowest task.

These failures arise because conventional agents treat plans, progress, evidence, and failures as transient conversation content. Search state should be maintained by the system rather than inferred repeatedly from interaction history. A long-horizon search system should expose what entities and attributes have been covered, preserve the provenance of every value, coordinate workers around explicit gaps, and intervene when execution stalls. It should also retain effective search and access patterns across sessions instead of rediscovering them for every task.

Based on this principle, we present SearchOS, a multi-agent framework for structured and observable open-domain information seeking. We first formulate an information-seeking task as relational schema completion with grounded citations. Following relational modeling and normalization principles (Codd, 1970; Bernstein, 1976), a relational search schema may contain multiple tables connected through primary–foreign-key relations. The system jointly discovers the entity rows, fills their attributes, and maintains a citation matrix that maps each value to a source URL and an anchored excerpt. This formulation, related to recent table-completion views of long-horizon information seeking (Lan et al., 2026), turns an open-ended request into a concrete search target and makes progress measurable at the level of entities and attributes.

We then introduce Search-Oriented Context Management (SOCM) to externalize execution state and share Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory across all agents. A central orchestrator decomposes the schema into coverage-oriented tasks and coordinates specialized explore, search, and writer agents. Rather than dispatching workers in synchronized batches, SearchOS uses pipeline-parallel scheduling: different sub-tasks progress concurrently, and each released agent slot is immediately assigned a new unresolved schema gap. This design follows the utilization advantage demonstrated by pipeline parallelism in GPU training (Huang et al., 2019; Narayanan et al., 2019); it reduces straggler-induced idle time and increases search throughput.

Reliable execution also requires controls that should not depend on an agent remembering to invoke them. We therefore introduce a Search Tool Middleware Harness that intercepts model and tool interactions to inject relevant state, extract and anchor evidence, enforce budgets, and detect repeated or stalled trajectories. Agents can focus on local search decisions while the harness maintains global execution invariants. Finally, we design a hierarchical search skill system that separates reusable strategy skills—how to search—from site-specific access skills—how to reach and extract information from a source. The system routes these skills by task and source and refines them from successful and failed trajectories.

Experiments on WideSearch and GISA show that SearchOS leads all reported F1 metrics among the evaluated single- and multi-agent baselines. It achieves 80.3 item-level F1 on WideSearch and 76.5 set F1 on GISA, exceeding the strongest baseline on the latter by 13.4 points, paving the way toward robust information-seeking collaboration. Our main contributions are as follows:

1. 1.

   Relational Search Formulation. We formulate open-domain information seeking as relational schema completion with grounded citations, providing a unified and verifiable objective for entity discovery, attribute completion, and evidence attribution.

2. 2.

   Stateful, Pipeline-Parallel Collaboration. We introduce SOCM and a continuous orchestration scheme that coordinate specialized agents through shared search state and dynamically assign work around unresolved coverage gaps.

3. 3.

   Middleware-Governed Search Agent Execution. We introduce a Search Tool Middleware Harness that moves evidence processing, context management, budget enforcement, and behavioral safeguards outside agent prompts.

4. 4.

   Hierarchical Search Skills. We design a hierarchical search agent skills system, comprising reusable search strategies and site-specific access procedures across sessions to augment the agents’ search process.

2 Problem Formulation

We formulate open-domain information seeking as relational schema completion with grounded citations. Following relational database design, we use primary keys to identify entities, foreign keys to express cross-table relations, and normalized tables to reduce redundant or inconsistent fact representations (Codd, 1970; Bernstein, 1976).

Given a natural-language request , the system constructs a relational search schema

(1)

where defines a table with attributes and primary-key attributes , and contains foreign-key relations between tables. The task is denoted by .

For each table, the system must discover a set of entities and populate a value matrix . Unlike conventional table completion, every populated value must be grounded in source evidence. We therefore maintain a citation matrix of the same shape, where each entry links a value to a source URL and an anchored excerpt:

(2)

The relational structure provides an explicit target for search, while the citation matrix makes each factual value independently verifiable. The completed schema is finally synthesized into a natural-language report with fine-grained inline citations.

3 SearchOS

To keep execution aligned with the evolving search state, we implement SearchOS as a stateful closed loop (Figure 2). We dispatch unresolved schema gaps through the orchestrator, commit observations through SOCM, and ground each interaction through middleware. Once complete, an optional writer produces the final report.

3.1 Search-Oriented Context Management

Because long-horizon search must track unresolved schema gaps, grounded evidence, coverage, and failed attempts, we design SOCM to externalize this information as a durable shared state

(3)

where is Frontier Task, the Evidence Graph, the Coverage Map, and Failure Memory. This state resides outside agent conversations and uses locked read–modify–write updates. An agent with role and assignment receives

(4)

rather than the full state. We use role-specific projections to expose global gaps to the orchestrator, assigned cells to search agents, and a compact summary to the writer. Each projection is regenerated from the latest state to avoid stale model context.

3.1.1 Frontier Task

To turn schema gaps into schedulable work, we maintain Frontier Task as a dependency-aware task pool. We represent each task and the ready set as

(5)

where , , , , , , and denote type, status, priority, dependencies, target cells, assigned agent, and attempt count; contains terminal task IDs. A task becomes pending after all dependencies terminate, then running on dispatch and completed on report collection; cycle, retry, or capacity guards may cancel it. Target-cell deduplication rejects overlapping active work, while dispatch-time revalidation closes tasks whose cells are already filled.

3.1.2 Evidence Graph

To preserve fine-grained provenance, we store atomic findings rather than page-level summaries in the Evidence Graph:

(6)

A node records value , source , supporting span , schema binding , confidence , provenance tier , and status . Rejected or superseded nodes remain for audit but do not count toward coverage. Deduplication uses binding, normalized value, and source, preserving corroborating sources. Provenance ranks anchored spans above unanchored pages and derived summaries.

3.1.3 Coverage Map

To make progress measurable, we materialize each schema cell in the Coverage Map with status (missing, filled, uncertain, or unreachable), supporting set , primary evidence, and a conflict flag. When several findings support a cell, we select

(7)

in lexicographic order, where is provenance tier, schema alignment, and authority-adjusted confidence. Disagreement marks a conflict rather than overwriting evidence.

Let be the materialized cells and the discovered rows of table . Coverage is computed as

(8)

The denominator keeps a declared empty table unfinished. In open-set tables, entity discovery expands , so coverage measures known rows rather than exhaustive enumeration.

3.1.4 Failure Memory

Multi-agent search requires a shared record of unsuccessful actions; otherwise, one agent may repeat a path that another has already exhausted. Failure Memory stores each record as

(9)

where is the failure type, its task-scoped signature, the corrective guidance, the recurrence count, and the latest occurrence. Recorded patterns include uninformative or repeated queries, inaccessible sources, failed skills, dropped branches, and rejected claims. Matching failures increment ; role-specific projections expose relevant records, and recurrence strengthens “do not retry” guidance. Post-mortems from failed runs can seed memory for related tasks.

3.1.5 Atomic State Updates

SOCM updates its four memories through the same locked state interface. Evidence-producing observations jointly update the Evidence Graph and Coverage Map, while agent reports and middleware sensors update Frontier Task and Failure Memory. This coupling prevents agents from observing evidence, coverage, task status, or failure guidance from different state versions; the corresponding middleware transitions are detailed in Sections 3.3.2 and 3.3.3.

3.2 Pipeline-Parallel Multi-Agent Orchestration

Agent roles.

To separate global orchestration from local search, we adopt an orchestrator–worker architecture (Wu et al., 2023). The orchestrator builds the schema, prioritizes gaps, and decides when to stop; explore agents identify candidates and sources; search agents collect grounded evidence for scoped gaps; and the writer produces a cited report. We coordinate these roles through SOCM rather than agent-to-agent conversation.

Continuous dispatch.

To avoid idle time from synchronized batches, we apply the utilization benefits of pipeline parallelism (Huang et al., 2019; Narayanan et al., 2019) to role-scoped work, synchronized through SOCM. At time , let be the number of available execution slots and the ready set defined by Frontier Task. We dispatch

(10)

where is frontier priority. After each completion, we update SOCM, recompute , and refill the released slot. This event-driven policy overlaps roles when dependencies permit and avoids batch stragglers.

Toolset design.

To keep each role’s action space focused, we expose role-scoped toolsets (Table 1). We reserve schema and task mutation for the orchestrator, browsing for explore and search agents, skill loading for search agents, and outline and state access for the writer. Browser observations update and only through middleware.

媒体内容 · 前往原文查看

Tool Group	Orchestrator	Search	Explore	Writer
Simple Browser (search/open/find)	–	✓	✓	–
Schema & Entity CRUD	✓	–	–	–
Task Queue & Coordination	✓	–	–	–
Outline Management	–	–	–	✓
SOCM Read	–	–	–	✓
Skill Catalog (list/load)	–	✓	–	✓

Browser interface.

The browser represents search as a shared navigation stack over three operations: search returns result pages, open renders and scrolls a selected source, and find locates keyword-matched spans. Each stack transition yields an observation; middleware determines evidence extraction and provenance. Appendix C gives the complete signatures.

3.3 Search Tool Middleware Harness

During long-horizon search, search agents face both model-level failures—such as losing task state or ignoring safeguards as context grows—and system-level disruptions, including tool errors, malformed outputs, stalled loops, and budget overruns. Because prompt-level safeguards cannot reliably cover heterogeneous post-trained agents, we introduce a system-level Search Tool Middleware Harness that intercepts the agent loop at model and tool boundaries. As illustrated in Figure 3, its three components form the following execution flow:

(11)

where is the role-specific model context, is a tool observation, contains runtime counters, and is a continue, correct, or stop action. Context is prepared before inference; after tool execution, evidence is grounded and committed before sensors evaluate progress and resource use.

3.3.1 Context Middleware

Before each model call, Context Middleware composes recent history , shared state , and retrieved skills for role :

				(12)

where projects shared state, is the selected skill set, renders skills, and composes context. The projection retains targets, grounded evidence, gaps, and failures without serializing the full history. Under budget pressure, older interactions are trimmed while state and active constraints remain.

3.3.2 Evidence Extraction Middleware

Because a visited page is not necessarily evidence, we use Evidence Extraction Middleware to extract schema-bound candidates from observation :

(13)

where each candidate contains an entity, attribute, value, source, and supporting span. Acceptance requires both schema binding and span anchoring:

(14)

We commit accepted evidence in one transition:

(15)

This atomically records provenance and updates Coverage Map cells, making evidence capture a consequence of tool execution.

3.3.3 Sensor Middleware

To detect stalls and enforce resource limits, we use Sensor Middleware to measure coverage and evidence progress. Let denote grounded schema coverage:

(16)

Over window , the trajectory stalls when neither quantity grows:

(17)

Budget pressure is the maximum normalized use of iterations, searches, and time:

(18)

Sensor Middleware computes the stall state and budget pressure . The harness combines these signals with current gaps to decide whether to continue execution, inject a correction, request backfill, enter drain-only mode, or stop a branch. Because the harness enforces these transitions outside prompts, the same safeguards apply across models and roles.

3.4 Hierarchical Search Skills

Because search requires reusable knowledge at different operational levels, we organize it into orchestrator, strategy, and access skills for global coordination, task-level methods, and source-specific retrieval. SearchOS-V1 contains 280 pre-built skills across domains (Figure 4).

Orchestrator skills.

Orchestrator skills are global playbooks for task decomposition, schema and row alignment, and synthesis validation; the current skill also enforces table and row-set alignment.

Strategy skills.

Strategy skills encode source-independent methods for query reformulation, entity enumeration and disambiguation, structured extraction, multi-hop and temporal reasoning, aggregation, and stall recovery.

Access skills.

Access skills encode source-specific retrieval and extraction for databases, government portals, encyclopedias, company sites, and media catalogs. Some combine instructions with typed executors to avoid repeated navigation and parsing; Appendix B illustrates one for Senate.gov.

Selection and execution.

At startup, we inject orchestrator skills as a shared playbook. For each task, we retrieve strategy and access skills: strategies become guidance, while executable access skills become typed tools. A query-driven router prunes the access catalog before selection.

4 Experiments

4.1 Benchmarks

We evaluate SearchOS on two open-domain information-seeking benchmarks as follows:

WideSearch (Wong et al., 2025). It targets large-scale “wide” information collection: it comprises 200 manually curated questions (100 English, 100 Chinese) drawn from real user queries across more than 15 domains, each requiring an agent to gather a large set of objectively verifiable atomic facts and organize them into a complete table.

GISA (Zhu et al., 2026). It evaluates general information-seeking assistants with 373 human-crafted queries reflecting authentic search scenarios; answers follow four structured formats (item, set, list, and table) that enable deterministic scoring, and the tasks couple deep multi-hop reasoning with broad cross-source aggregation.

4.2 Baselines

We compare SearchOS against typical baselines in two categories: single-agent methods, ReAct (Yao et al., 2022) and Plan-and-Solve (Wang et al., 2023), and multi-agent systems, A-MapReduce (Chen et al., 2026), Web2BigTable (Huang et al., 2026), and Table-as-Search (Lan et al., 2026).

4.3 Experimental Settings

Metrics. Following the two benchmarks, we score the assembled tables against the gold standard at two granularities. Item-level metrics (Precision / Recall / F1) treat each answer cell independently, while Row-level metrics (Precision / Recall / F1) credit a row only when all of its cells are correct. We additionally report Exact Match (EM), the fraction of cases whose entire table exactly matches the gold answer. On GISA we break out F1 by question type (Table / Set / List / Item).

Configuration. We use GLM-5 as the backbone for the agent roles and Qwen3.5-35B-A3B for Evidence Extraction. For each case we run the system three times and report the best of the three runs (Max@3), scaled by . Unless otherwise specified, we cap each session at 50 orchestrator iterations, 8 parallel sub-agents, 20 searches per sub-agent, and a 1800-second wall-clock budget.

4.4 Main Results

媒体内容 · 前往原文查看

Single Agent Multi-Agent System Benchmark Metric ReAct Plan-and-Solve Table-as-Search A-MapReduce Web2BigTable Ours WideSearch Item Precision 82.9 83.8 82.4 83.1 78.3 83.9 +0.1 Item Recall 70.2 72.9 73.5 74.2 73.4 79.7 +5.5 Item F1 72.9 75.2 75.4 76.0 73.8 80.3 +4.3 Row Precision 58.0 58.7 57.1 56.9 57.5 59.0 +0.3 Row Recall 48.8 50.2 51.6 49.8 54.0 55.8 +1.8 Row F1 50.9 52.2 52.7 51.4 54.5 56.5 +2.0 GISA Table Item F1 74.8 71.2 73.4 72.5 68.1 76.9 +2.1 Table Row F1 58.1 50.7 54.1 52.1 45.3 59.7 +1.6 Set F1 61.6 63.1 60.9 62.5 56.7 76.5 +13.4 List F1 67.1 53.8 54.2 57.4 65.5 68.1 +1.0 Item EM 0.0 16.7 16.7 33.3 50.0 50.0 0.0

Table 2 reports the main results. SearchOS achieves the best overall performance on both benchmarks, leading on all six headline F1 metrics across the two benchmarks while remaining competitive on the precision-oriented metrics.

On WideSearch, SearchOS attains the best Item-level Precision (), Recall (), and F1 (), improving over the strongest baseline (A-MapReduce, F1) by points. The gain is concentrated in recall, consistent with the design goal of using coverage-aware dispatch to reduce missing cells, while precision remains the highest among all methods. On the stricter Row-level metrics, where a single wrong cell invalidates the entire row, SearchOS also leads in Precision (), Recall (), and F1 (, over the next-best Web2BigTable). Thus, the recall advantage extends to full-row consistency without sacrificing precision.

On GISA, SearchOS leads across every question type. The gains are largest on Set questions ( vs. the best baseline’s , a -point improvement), where correctly enumerating a complete answer set benefits directly from the Evidence Graph and coverage-driven completeness checks. SearchOS also tops Table F1 at both the Item () and Row () levels and List F1 (), and matches the best Item-level EM (). Overall, the results show that formulating open-domain information seeking as relational schema completion, backed by shared SOCM state, yields the most consistent gains precisely on the recall- and completeness-sensitive metrics that long-horizon search tasks demand.

5 Ablations & Analysis

In this section, we present some insightful ablation and analytical experiments to further evaluate the performance of SearchOS.

5.1 Fixed Schemas vs. Search-Time Schema Planning

We first test whether fixed schema structure suits all tasks. We first select 40 questions that can be decoupled into multi-table schema. For each question, we use GPT-5.5 to construct semantically matched fixed single-table and multi-table schemas containing the same required attributes. Then we run SearchOS under each fixed schema and compare them with our native autonomous schema planning with explore agent.

Table 3 reports the comparison on 40 cases. Fixed multi-table is stronger on average than fixed single-table, but wins on only 21 cases, loses on 17, and ties on two by Item F1, showing that their relative advantage is task-dependent. Even an oracle that selects the better fixed schema for each case remains below SearchOS by Item F1 and Row F1 points.

媒体内容 · 前往原文查看

Item-level Row-level Setting Precision Recall F1 Precision Recall F1 Fixed Single-Table 54.7 46.7 47.9 33.5 28.2 29.2 Fixed Multi-Table 66.9_(+12.2) 55.7_(+9.0) 58.3_(+10.4) 37.4_(+3.9) 32.6_(+4.4) 34.2_(+5.0) Oracle Single/Multi 70.3_(+3.4) 60.7_(+5.0) 62.4_(+4.1) 44.8_(+7.4) 39.7_(+7.1) 41.2_(+7.0) SearchOS 76.3_(+6.0) 68.3_(+7.6) 70.6_(+8.2) 52.7_(+7.9) 47.3_(+7.6) 48.9_(+7.7)

SearchOS Schema Choice

Single 87.5%   Multi 12.5%

SearchOS dynamically selects the task-appropriate schema.

Our experiments show that different tasks favor different table structures. Tasks with concentrated entity attributes and simple relations may be better represented by a single table, whereas those involving one-to-many relations, multiple entity types, shared attributes, or cross-entity organization are more likely to benefit from multiple tables. Across the 40 cases, SearchOS selected a single-table schema for 35 tasks and a multi-table schema for five. Neither fixed structure is universally optimal.

The consistent advantage of SearchOS shows that schema planning should remain part of the search process rather than be fixed in advance. By retaining the complete search-and-planning loop, the system can organize information according to the entity types, field distributions, repetition patterns, and relations encountered during search, choosing a single-table, multi-table, or hybrid structure as appropriate and revising it when new evidence changes the information topology.

5.2 Pipeline-Parallel Scheduling Ablation

媒体内容 · 前往原文查看

Round Time () Tokens () Item F1 (pp) 1 % % 2 % % 3 % %

We next evaluate whether pipeline parallel continuous dispatch improves pipeline utilization over synchronized batches. We conduct three paired rounds with the same queries, model configuration, and concurrency limit (), changing only the scheduling policy on 10 WideSearch cases. Batch scheduling waits for all tasks in a group to finish, whereas continuous scheduling immediately refills each released slot.

Table 4 reports the paired results in each round. Continuous scheduling consistently reduces end-to-end time and token use while improving Item F1, showing that its advantage is stable across repeated runs.

媒体内容 · 前往原文查看

Scheduling Policy Time (s) () Slot Utilization () Tasks/min () LLM Calls () Item F1 () Batch (Control) 629.13 34.6% 2.99 341.4 79.66 Continuous (Ours) 476.34 41.7% 3.37 296.6 86.75

Continuous scheduling reduces average end-to-end time by and improves slot utilization and task throughput, while using fewer LLM calls and achieving higher Item F1. These results show that continuous dispatch removes idle capacity at batch barriers and improves efficiency without sacrificing retrieval quality.

5.3 Middleware-Governed Search Agent Execution Analysis

Beyond aggregate scores, we further analyze whether our Search Tool Middleware Harness governance keeps long-horizon search productive after detecting stagnation.

Figure 5 presents three representative WideSearch trajectories with early, mid-run, and late interventions. In each column, the upper figure shows table coverage and the lower figure shows the cumulative number of discovered entities. The dashed line marks the point at which the Loop Sensor detects a low-progress search loop and triggers a strategy switch.

Across the different cases, interventions at early, middle, and late stages all restore search progress beyond the preceding stagnation period, although they affect coverage and entity discovery to different degrees. These trajectories provide mechanism-level evidence that the Loop Sensor can redirect stalled search toward productive work.

5.4 Skill Ablation and Efficiency

We compare SearchOS with and without hierarchical skills to assess their contribution to search quality and efficiency. Table 6 reports Max@3 over three runs per setting.

媒体内容 · 前往原文查看

Item-level Row-level Setting Precision Recall F1 Precision Recall F1 Without Skills 82.2 78.5 78.3 56.0 52.4 53.1 With Skills 83.9 79.7 80.3 59.0 55.8 56.5 +1.7 +1.2 +2.0 +3.0 +3.4 +3.4

Effectiveness.

Skills raise Item F1 by points and Row F1 by points. The larger row-level gain suggests that reusable decomposition, search, and source-access knowledge helps agents assemble coherent entity records rather than merely recover isolated cells.

媒体内容 · 前往原文查看

Efficiency.

Figure 6 shows that skills reduce session time by , search calls by , and page calls by . The quality gains therefore come with less exploratory trial and error rather than additional browsing. Since all skill layers are disabled together, this ablation measures their joint contribution.

6 Related Work

6.1 Agentic Information Seeking

Tool-integrated language agents interleave search, browsing, reasoning, evidence collection, and synthesis (Nakano et al., 2021; Yao et al., 2022; Trivedi et al., 2023; Qin et al., 2023; Chen et al., 2025b; Jin et al., 2025a; Li et al., 2026b; Zheng et al., 2025; Li et al., 2025). Their capabilities have advanced through search policy optimization (Qian and Liu, 2025; Luo et al., 2025), task decomposition (Chen et al., 2025b, 2026), synthetic interaction data (Tao et al., 2025; Li et al., 2025), and reinforcement-based post-training (Wang et al., 2025b; Jin et al., 2025a; Gao et al., 2025; Chen et al., 2025a; Zheng et al., 2025). Yet progress typically remains in growing interaction histories, obscuring established evidence and unresolved questions.

Broad information seeking has also been formulated as structured or table-oriented completion (Tao et al., 2025; Wong et al., 2025; Zhu et al., 2026; Lan et al., 2026; Chen et al., 2026; Huang et al., 2026). Such outputs expose missing entities and attributes and support parallel collection, but do not capture task dependencies, conflicting observations, provenance, or failed attempts. SearchOS instead formulates the task as relational schema completion with grounded citations and maintains these intermediate artifacts outside the interaction history.

6.2 Multi-Agent Orchestration Systems

Multi-agent frameworks decompose complex tasks through specialized roles, message exchange, and structured workflows (Wu et al., 2023; Hong et al., 2024; Chen et al., 2024; Qian et al., 2024; Chen et al., 2025b; Fourney et al., 2024); hierarchical approaches decouple strategic planning from specialized execution (Jin et al., 2025b), while structured execution systems distinguish dependent from parallel sub-tasks and execute independent branches concurrently (Kim et al., 2024; Zhang et al., 2025). However, coordination state usually remains in conversations, plans, or general task ledgers, making redundant collection, stale work, coverage gaps, and idle capacity difficult to manage explicitly during long-horizon search.

Recent advances in agent memory and context management address related challenges through reflection, memory stores, virtual context, compression, or trajectory consolidation (Park et al., 2023; Shinn et al., 2023; Zhong et al., 2024; Packer et al., 2023; Jiang et al., 2023; Ye et al., 2025). In contrast, Search-Oriented Context Management (SOCM) externalizes intermediate search state as a Frontier Task, Evidence Graph, Coverage Map, and Failure Memory, with role-specific projections for each agent.

6.3 Agent Harness Engineering

Recent surveys treat the agent harness as a first-class infrastructure layer governing execution, tools, context, state, lifecycle, observability, and evaluation, while engineering reports emphasize inspectable environments, feedback loops, and context curation for reliable long-horizon agents (Li et al., 2026a; Meng et al., 2026; Lopopolo, 2026; Anthropic, 2025). Prior systems instantiate parts of this layer through environment middleware, message-driven and fault-tolerant runtimes, operating-system abstractions, efficient model–tool interception, and runtime constraints (Gu et al., 2024; Gao et al., 2024; Mei et al., 2025; Abhyankar et al., 2024; Wang et al., 2025a; Ning et al., 2026; Zhang et al., 2026). The SearchOS Search Tool Middleware Harness applies these ideas to long-horizon information seeking by preparing role-specific context, validating and grounding evidence-producing observations, and monitoring repetition, stalls, and budget exhaustion.

7 Conclusion

In this work, we introduce SearchOS, a multi-agent framework that externalizes long-horizon open-domain search as system-maintained state. We formulate requests as relational schema completion with grounded citations and use Search-Oriented Context Management (SOCM) to organize coverage gaps, evidence, and intermediate search artifacts. Pipeline-parallel orchestration dispatches unresolved gaps, while our Search Tool Middleware Harness prepares context, grounds observations, and handles stalls and budget limits. A hierarchical skill library supplies reusable orchestration, search-strategy, and source-access knowledge.

Experiments on WideSearch and GISA show that SearchOS improves both completeness and efficiency. These results suggest that explicit search state, grounded evidence, and system-level execution control provide a practical foundation for reliable multi-agent information seeking. We will extend SearchOS to broader domains and multimodal settings while improving adaptation across agents, sources, and tasks.

References

• Abhyankar et al. (2024) R. Abhyankar, Z. He, V. Srivatsa, H. Zhang, and Y. Zhang. InferCept: Efficient intercept support for augmented large language model inference. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 81–95. PMLR, 2024. URL https://proceedings.mlr.press/v235/abhyankar24a.html.
• Anthropic (2025) Anthropic. Effective context engineering for AI agents. Anthropic Engineering, Sept. 2025. URL https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents.
• Bernstein (1976) P. A. Bernstein. Synthesizing third normal form relations from functional dependencies. ACM Transactions on Database Systems (TODS), 1(4):277–298, 1976.
• Chen et al. (2025a) M. Chen, L. Sun, T. Li, H. Sun, Y. Zhou, C. Zhu, H. Wang, J. Z. Pan, W. Zhang, H. Chen, F. Yang, Z. Zhou, and W. Chen. Research: Learning to reason with search for llms via reinforcement learning, 2025a. URL https://arxiv.org/abs/2503.19470.
• Chen et al. (2026) M. Chen, G. Zhang, H. Chang, Y. Guo, and S. Zhou. A-mapreduce: Executing wide search via agentic mapreduce. arXiv preprint arXiv:2602.01331, 2026.
• Chen et al. (2024) W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C.-M. Chan, H. Yu, Y. Lu, Y.-H. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=EHg5GDnyq1.
• Chen et al. (2025b) Z. Chen, K. Liu, Q. Wang, J. Liu, W. Zhang, K. Chen, and F. Zhao. MindSearch: Mimicking human minds elicits deep ai searcher. In International Conference on Learning Representations, 2025b. URL https://openreview.net/forum?id=xgtXkyqw1f.
• Codd (1970) E. F. Codd. A relational model of data for large shared data banks. Communications of the ACM, 13(6):377–387, 1970.
• Fourney et al. (2024) A. Fourney, G. Bansal, H. Mozannar, C. Tan, E. Salinas, E. Zhu, F. Niedtner, G. Proebsting, G. Bassman, J. Gerrits, J. Alber, P. Chang, R. Loynd, R. West, V. Dibia, A. Awadallah, E. Kamar, R. Hosn, and S. Amershi. Magentic-One: A generalist multi-agent system for solving complex tasks. arXiv preprint arXiv:2411.04468, 2024. URL https://arxiv.org/abs/2411.04468.
• Gao et al. (2024) D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y. Xie, D. Chen, L. Yao, H. Peng, Z. Zhang, L. Zhu, C. Cheng, H. Shi, Y. Li, B. Ding, and J. Zhou. AgentScope: A flexible yet robust multi-agent platform. arXiv preprint arXiv:2402.14034, 2024. URL https://arxiv.org/abs/2402.14034.
• Gao et al. (2025) J. Gao, W. Fu, M. Xie, S. Xu, C. He, Z. Mei, B. Zhu, and Y. Wu. Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl, 2025. URL https://arxiv.org/abs/2508.07976.
• Gu et al. (2024) Y. Gu, Y. Shu, H. Yu, X. Liu, Y. Dong, J. Tang, J. Srinivasa, H. Latapie, and Y. Su. Middleware for LLMs: Tools are instrumental for language agents in complex environments. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7646–7663, 2024. 10.18653/v1/2024.emnlp-main.436. URL https://aclanthology.org/2024.emnlp-main.436/.
• Hong et al. (2024) S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, L. Zhou, C. Ran, L. Xiao, C. Wu, and J. Schmidhuber. MetaGPT: Meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=VtmBAGCN7o.
• Huang et al. (2019) Y. Huang, Y. Cheng, A. Bapna, O. Firat, M. X. Chen, D. Chen, H. Lee, J. Ngiam, Q. V. Le, Y. Wu, and Z. Chen. GPipe: Efficient training of giant neural networks using pipeline parallelism. In Advances in Neural Information Processing Systems, volume 32, pages 103–112, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/093f65e080a295f8076b1c5722a46aa2-Abstract.html.
• Huang et al. (2026) Y. Huang, Y. Chen, Z. He, Y. Chen, K. Y. Lee, H. Zhou, W. Luo, M. Fang, and J. Wang. Web2bigtable: A bi-level multi-agent llm system for internet-scale information search and extraction. arXiv preprint arXiv:2604.27221, 2026.
• Jiang et al. (2023) H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y. Lin, Y. Yang, and L. Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. arXiv preprint arXiv:2310.06839, 2023.
• Jin et al. (2025a) B. Jin, H. Zeng, Z. Yue, D. Wang, H. Zamani, and J. Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025a. URL https://arxiv.org/abs/2503.09516.
• Jin et al. (2025b) J. Jin, X. Li, G. Dong, Y. Zhang, Y. Zhu, Y. Zhao, H. Qian, and Z. Dou. Hira: A hierarchical reasoning framework for decoupled planning and execution in deep search. arXiv preprint arXiv:2507.02652, 2025b.
• Kim et al. (2024) S. Kim, S. Moon, R. Tabrizi, N. Lee, M. W. Mahoney, K. Keutzer, and A. Gholami. An LLM compiler for parallel function calling. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 24370–24391. PMLR, 2024. URL https://proceedings.mlr.press/v235/kim24y.html.
• Lan et al. (2026) T. Lan, F. Henry, B. Zhu, Q. Jia, J. Ren, Q. Pu, H. Li, L. Wang, Z. Xu, and W. Luo. Table-as-search: Formulate long-horizon agentic information seeking as table completion. arXiv preprint arXiv:2602.06724, 2026.
• Li et al. (2026a) J. Li, X. Xiao, Y. Zhang, C. Liu, L. Zhao, X. Liao, Y. Ji, J. Wang, J. Gu, Y. Ge, W. Xu, X. Fang, X. Xu, T. Zhao, Y. Kim, T. Wang, J. Hamm, S. Krishnaswamy, J. Huan, and C. K. Reddy. Agent harness engineering: A survey. OpenReview preprint, 2026a. URL https://openreview.net/forum?id=eONq7FdiHa.
• Li et al. (2025) K. Li, Z. Zhang, H. Yin, R. Ye, Y. Zhao, L. Zhang, L. Ou, D. Zhang, X. Wu, J. Wu, X. Wang, Z. Qiao, Z. Zhang, Y. Jiang, P. Xie, F. Huang, and J. Zhou. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning. CoRR, abs/2509.13305, 2025. 10.48550/ARXIV.2509.13305. URL https://doi.org/10.48550/arXiv.2509.13305.
• Li et al. (2026b) X. Li, J. Jin, G. Dong, H. Qian, Y. Wu, J.-R. Wen, Y. Zhu, and Z. Dou. Webthinker: Empowering large reasoning models with deep research capability. Advances in Neural Information Processing Systems, 38:120091–120131, 2026b.
• Lopopolo (2026) R. Lopopolo. Harness engineering: Leveraging Codex in an agent-first world. OpenAI Engineering, Feb. 2026. URL https://openai.com/index/harness-engineering/.
• Luo et al. (2025) K. Luo, H. Qian, Z. Liu, Z. Xia, S. Xiao, S. Bao, J. Zhao, and K. Liu. Infoflow: Reinforcing search agent via reward density optimization, 2025. URL https://arxiv.org/abs/2510.26575.
• Mei et al. (2025) K. Mei, X. Zhu, W. Xu, M. Jin, W. Hua, Z. Li, S. Xu, R. Ye, Y. Ge, and Y. Zhang. AIOS: LLM agent operating system. In Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=L4HHkCDz2x.
• Meng et al. (2026) Q. Meng, Y. Wang, L. Chen, Y. Li, W. Wu, W. Jiang, Q. Wang, C. Lu, Y. Gao, Y. Wu, and Y. Hu. Agent harness for large language model agents: A survey. Preprints, 2026. 10.20944/preprints202604.0428.v3. URL https://www.preprints.org/manuscript/202604.0428.
• Nakano et al. (2021) R. Nakano, J. Hilton, S. Balaji, J. Wu, L. Ouyang, C. Kim, C. Hesse, S. Jain, V. Kosaraju, W. Saunders, X. Jiang, K. Cobbe, T. Eloundou, G. Krueger, K. Button, M. Knight, B. Chess, and J. Schulman. Webgpt: Browser-assisted question-answering with human feedback. CoRR, abs/2112.09332, 2021. URL https://arxiv.org/abs/2112.09332.
• Narayanan et al. (2019) D. Narayanan, A. Harlap, A. Phanishayee, V. Seshadri, N. R. Devanur, G. R. Ganger, P. B. Gibbons, and M. Zaharia. PipeDream: Generalized pipeline parallelism for DNN training. In Proceedings of the 27th ACM Symposium on Operating Systems Principles, pages 1–15, 2019. 10.1145/3341301.3359646. URL https://doi.org/10.1145/3341301.3359646.
• Ning et al. (2026) X. Ning, K. Tieu, D. Fu, T. Wei, Z. Li, Y. Bei, J. Zou, M. Ai, Z. Liu, T.-W. Li, L. Chen, Y. Zhao, K. Yang, B. Li, C. Qian, G. Li, X. Lin, Z. Zeng, R. Qiu, S. Chen, Y. Sun, X. Yang, R. Wang, R. Pan, C. Yang, D. Zhang, L. Fang, Z. Cui, Y. Cao, P. Chen, D. Sun, R. Chen, M. Srinivasan, N. Mathur, Y. Xia, H. Li, H. Yan, P. Lu, L. Zhang, T. Zhang, H. Tong, and J. He. Code as agent harness. arXiv preprint arXiv:2605.18747, 2026. URL https://arxiv.org/abs/2605.18747.
• Packer et al. (2023) C. Packer, S. Wooders, K. Lin, V. Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez. MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560.
• Park et al. (2023) J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023. 10.1145/3586183.3606763. URL https://doi.org/10.1145/3586183.3606763.
• Qian et al. (2024) C. Qian, W. Liu, H. Liu, N. Chen, Y. Dang, J. Li, C. Yang, W. Chen, Y. Su, X. Cong, J. Xu, D. Li, Z. Liu, and M. Sun. ChatDev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174–15186, 2024. 10.18653/v1/2024.acl-long.810. URL https://aclanthology.org/2024.acl-long.810/.
• Qian and Liu (2025) H. Qian and Z. Liu. Scent of knowledge: Optimizing search-enhanced reasoning with information foraging. arXiv preprint arXiv:2505.09316, 2025.
• Qin et al. (2023) Y. Qin, Z. Cai, D. Jin, L. Yan, S. Liang, K. Zhu, Y. Lin, X. Han, N. Ding, H. Wang, et al. Webcpm: Interactive web search for chinese long-form question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8968–8988, 2023.
• Shinn et al. (2023) N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html.
• Tao et al. (2025) Z. Tao, J. Wu, W. Yin, J. Zhang, B. Li, H. Shen, K. Li, L. Zhang, X. Wang, Y. Jiang, P. Xie, F. Huang, and J. Zhou. Webshaper: Agentically data synthesizing via information-seeking formalization. CoRR, abs/2507.15061, 2025. 10.48550/ARXIV.2507.15061. URL https://doi.org/10.48550/arXiv.2507.15061.
• Trivedi et al. (2023) H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pages 10014–10037, 2023.
• Wang et al. (2025a) H. Wang, C. M. Poskitt, and J. Sun. AgentSpec: Customizable runtime enforcement for safe and reliable LLM agents. arXiv preprint arXiv:2503.18666, 2025a. URL https://arxiv.org/abs/2503.18666.
• Wang et al. (2023) L. Wang, W. Xu, Y. Lan, Z. Hu, Y. Lan, R. K.-W. Lee, and E.-P. Lim. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023.
• Wang et al. (2025b) Z. Wang, X. Zheng, K. An, C. Ouyang, J. Cai, Y. Wang, and Y. Wu. Stepsearch: Igniting llms search ability via step-wise proximal policy optimization, 2025b. URL https://arxiv.org/abs/2505.15107.
• Wong et al. (2025) R. Wong, J. Wang, J. Zhao, L. Chen, Y. Gao, L. Zhang, X. Zhou, Z. Wang, K. Xiang, G. Zhang, W. Huang, Y. Wang, and K. Wang. Widesearch: Benchmarking agentic broad info-seeking. arXiv preprint arXiv:2508.07999, 2025.
• Wu et al. (2023) Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang. Autogen: Enabling next-gen LLM applications via multi-agent conversation. CoRR, abs/2308.08155, 2023. 10.48550/ARXIV.2308.08155. URL https://doi.org/10.48550/arXiv.2308.08155.
• Yao et al. (2022) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022.
• Ye et al. (2025) R. Ye, Z. Zhang, K. Li, H. Yin, Z. Tao, Y. Zhao, L. Su, L. Zhang, Z. Qiao, X. Wang, P. Xie, F. Huang, S. Chen, J. Zhou, and Y. Jiang. Agentfold: Long-horizon web agents with proactive context management, 2025. URL https://arxiv.org/abs/2510.24699.
• Zhang et al. (2025) Y. Zhang, Z. Dou, X. Li, J. Jin, Y. Wu, Z. Li, Y. Qi, and J. Wen. Neuro-symbolic query compiler. In W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, editors, Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025, pages 12138–12155. Association for Computational Linguistics, 2025. URL https://aclanthology.org/2025.findings-acl.628/.
• Zhang et al. (2026) Y. Zhang, H. Lu, J. Jin, H. Qian, S. Li, Z. Yang, Y. Zhu, J.-R. Wen, and Z. Dou. Web sitemap knowledge can enhance autonomous browsing. In Findings of the Association for Computational Linguistics: ACL 2026, pages 29307–29321, 2026.
• Zheng et al. (2025) Y. Zheng, D. Fu, X. Hu, X. Cai, L. Ye, P. Lu, and P. Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 414–431, 2025.
• Zhong et al. (2024) W. Zhong, L. Guo, Q. Gao, H. Ye, and Y. Wang. MemoryBank: Enhancing large language models with long-term memory. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724–19731, 2024. 10.1609/aaai.v38i17.29946.
• Zhu et al. (2026) Y. Zhu, X. Zhang, M. Zhang, J. Jin, L. Zhang, X. Song, K. Zhao, W. Zeng, R. Tang, H. Li, J.-R. Wen, and Z. Dou. Gisa: A benchmark for general information-seeking assistant. arXiv preprint arXiv:2602.08543, 2026.

Appendix A SOCM State Example

This appendix provides a concrete example of the SOCM state as injected into the orchestrator’s prompt during a search session.

Appendix B Executable Access Skill Case

An executable access skill is packaged as three aligned artifacts: skill.md provides source-specific usage guidance, manifest.yaml declares the typed parameter schema exposed to the agent, and executor.py implements dispatch, retrieval, parsing, and normalization. The following abridged case is taken from the current Senate.gov access skill.

The manifest constrains the agent to a small operation vocabulary, while the executor hides source-specific URL construction, HTTP error handling, and BeautifulSoup parsing. Consequently, the agent receives stable structured records even though the underlying evidence is extracted from heterogeneous Senate state-profile pages.

Appendix C Implementation Reference

For reproducibility, this appendix records the tool interfaces summarized in Section 3.2: the Simple Browser operations and the orchestrator’s schema, task-queue, and writer tools. These interfaces are implementation choices rather than part of the core method.

C.1 Simple Browser

Browser State Abstraction.

We model the browser state as a tuple , where is a page cache mapping URLs to rendered page contents, is a LIFO page stack recording the navigation history, and is a scroll cursor tracking the current viewport position within the top-of-stack page. Every browser operation that produces a new page—a search query, a page fetch, or a find operation—pushes the result onto , establishing a natural navigation history.

This stack-based design provides several advantages over a flat URL-based model:

• •

  Contextual link resolution: When the agent issues open(3), the browser walks the stack backwards to find the most recent page containing link ID 3, resolving the numeric reference unambiguously.

• •

  Implicit back-navigation: The stack preserves the full navigation path, enabling further trajectory analysis.

• •

  Cross-page find: find results reference the source page (the top non-find page on the stack), and subsequent open(match_id) calls auto-jump to the matched line, enabling efficient in-page exploration without losing context.

Browser Tools.

Table 7 summarizes the three browser operations.

媒体内容 · 前往原文查看

Tool	Parameters	Description
search	query: str	Web search; return URLs, titles, and snippets.
open	id_or_url: intstr; loc: int	Visit a page by URL or link ID, starting at a given line.
find	pattern: strlist	Grep keywords from the currently open page.

Figure 7 illustrates the rendered outputs an agent receives from open() and find() calls.

媒体内容 · 前往原文查看

C.2 Schema and Entity Management Tools

The orchestrator manages the relational search schema through a suite of CRUD tools that operate on the Coverage Map (Section 3.1.3). Table 8 lists the full tool set.

媒体内容 · 前往原文查看

Tool	Key Parameters	Description
create_schema	tables, relations	Define schema with tables, attributes, PKs, and FKs.
add_entities	entities, attributes	Add rows with optional inline values (stored as evidence).
remove_entities	entities, reason	Soft-delete rows with mandatory justification.
edit_entities	edits	Rename PKs or edit cells. Preserves evidence refs.
inspect_table	table_id, status	Return row list by status (filled/partial/empty).

C.3 Task Queue and Agent Coordination Tools

The orchestrator manages the Frontier Memory (Section 3.1.1) and sub-agent lifecycle through the tools listed in Table 9.

媒体内容 · 前往原文查看

Tool	Key Parameters	Description
enqueue_tasks	items	Append tasks with priority, target table, and dependencies.
check_agents	timeout: float	Block until 1 agent completes; return reports.
stop_task	task_ids, reason	Cancel tasks and their running sub-agents.

C.4 Writer Tools

The writer agent operates on a structured outline rather than producing free-form text, ensuring consistent organization and complete citation coverage. Table 10 lists the writer-specific tools.

媒体内容 · 前往原文查看

Tool	Key Parameters	Description
read_outline	section_id, toc_only	Return TOC or full section content.
update_outline	ops (JSON array)	Batch add/remove/reorder sections.
write_section	section_id, cited_evidence_ids	Write section with mandatory citations.
edit_section	section_id, old, new	Exact substring replacement for surgical edits.
annotate_section	section_id, annotation	Attach annotation to a section.

In addition, the writer agent has read-only access to the full SOCM state and the skill catalog, enabling it to ground its writing in the latest evidence.

Appendix D Agent Trajectory Case Studies

This section presents process-level case studies from complete WideSearch trajectories. We summarize the logged decision rationale, tool action, and observable state transition at each consequential step rather than reproducing unabridged token-level reasoning. The cases were selected to expose different behaviors—parallel enrichment, scope auditing, and recovery from inaccessible sources—and are illustrative rather than aggregate evidence.

D.1 Detailed Case: Spotify 2024 Rankings

Task and outcome. Case ws_en_022 asks for the global and U.S. top-ten songs on Spotify in 2024, enriched with artist, language, songwriter, producer, and release date. The run finishes in 362 seconds with 221 evidence nodes, 100% known-cell coverage, 97.5 Item F1, and 85.0 Row F1.

Decision trace.

Table 11 reconstructs the consequential intermediate decisions from the stored trajectory. The key design choice is to establish the two authoritative ranked lists before dispatching metadata enrichment, so each sub-agent writes into a stable composite key rather than independently rediscovering row identities.

媒体内容 · 前往原文查看

Stage	Decision rationale	Action	Observable state
1. Explore	First identify an authoritative source that fixes both top-ten lists.	Dispatch one Explore Agent to inspect Spotify Wrapped and chart sources.	No schema; coverage 0%.
2. Structure	The official lists expose two row dimensions: chart category and rank.	Create one table with primary key [Category, Rank]; seed both lists.	40/160 cells known; coverage 25%.
3. Parallelize	Metadata lookup is independent after row identities are fixed.	Split the 20 songs into four five-song tasks; dispatch four Search Agents with the Spotify access skill.	Four agents run concurrently.
4. Monitor	Merge completed shards without waiting for the slowest agent.	Repeatedly call check_agents; retain shared Coverage Map updates.	Coverage 25.068.889.4100.0%.
5. Audit	Full known-cell coverage does not guarantee formatting consistency or scope completion.	Inspect all 20 rows; compare known row set to the two official top-ten lists.	20 expected rows present; zero missing rows.
6. Repair	Two release dates use inconsistent separators.	Apply two targeted edit_entities operations, then export.	20 complete rows; 221 evidence nodes.

SOCM evolution.

The orchestrator receives compact state snapshots while the Search Agents work. Early in the run, all 20 seeded rows have gaps. A mid-run snapshot reports only six rows with missing or uncertain columns. The final snapshot reports zero gaps, while explicitly warning that “all known rows filled” is not proof that the requested row set is exhaustive. This warning triggers the Stage-5 row-set audit rather than immediate termination.

Output excerpt.

The same song may occur in both charts, but the composite key preserves its category-specific rank while shared metadata is independently grounded.

媒体内容 · 前往原文查看

List	Rank	Title	Singer	Release Date
Global	1	Espresso	Sabrina Carpenter	2024/04/11
Global	2	Beautiful Things	Benson Boone	2024/01/19
U.S.	1	Espresso	Sabrina Carpenter	2024/04/11
U.S.	2	Not Like Us	Kendrick Lamar	2024/05/04

D.2 Additional Process Cases

False saturation followed by scope repair (ws_en_016).

The task enumerates Michael Phelps’s individual-event medals across the Olympic Games and World Aquatics Championships. The orchestrator initially splits the work by competition family. After both agents return, known-cell coverage reaches 100%, but the table contains only 16 rows while the agent reports imply roughly 31 eligible events. Instead of treating saturation as completion, the orchestrator inspects the row set, identifies missing years and competitions, and dispatches targeted backfill tasks. The table grows to 33 and then 35 rows. A final verification pass detects a duplicate 2003 World Championships 400m individual-medley record, removes the incorrect 4:11.04 row, adds the missing 2005 100m butterfly result, and corrects medalists in several 200m freestyle rows. This trajectory illustrates why Coverage Map saturation is paired with explicit scope auditing.

Blocked publisher pages and bounded recovery (ws_en_023).

This task enumerates migration-related articles from three journals over 2020–2024. The orchestrator first corrects an ambiguous journal name, then assigns one Search Agent per journal. After the initial pass, it removes an out-of-scope 2026 article and dispatches targeted metadata tasks for five incomplete rows. Wiley and JSTOR pages repeatedly return access checks or HTTP 402 responses. The affected agents pivot to Crossref, OpenAlex, and academic-paper skills; the Loop Sensor records two explicit strategy switches for one stalled agent. When the remaining sources and per-agent budgets are exhausted, the system preserves the unresolved cells instead of fabricating metadata. Despite these access failures, the run returns 57 rows with 96.3% coverage and 93.8 Item and Row F1.

媒体内容 · 前往原文查看

Case Primary behavior Tasks Evidence Item F1 Row F1 ws_en_022 Parallel enrichment + final audit 5 221 97.5 85.0 ws_en_016 False-saturation repair + verification 8 486 80.1 28.2 ws_en_023 Source fallback + bounded completion 11 302 93.8 93.8

Cross-case observations.

The three traces expose complementary aspects of the system that are hidden by final-answer scores:

• •

  Explore before committing to row identities. In the Spotify case, an authoritative list establishes the complete key space before metadata agents are launched. This prevents independently discovered song records from drifting across the two chart categories.

• •

  Cell saturation and scope completion are separate conditions. In the Phelps case, 100% coverage over 16 known rows is correctly rejected as completion because the expected event count is much larger. Coverage controls attribute filling; an explicit row-set audit controls recall.

• •

  Recovery remains bounded and evidence preserving. In the journal case, sensors encourage source and tool changes after repeated failures, while resource limits eventually stop unproductive search. Unresolved cells remain visible instead of being filled from unsupported inference.

Together, these cases show the intended control loop: exploration defines a defensible search space, parallel agents accumulate evidence into shared structured state, SOCM exposes concrete residual gaps, and governance redirects or stops work according to observable progress.

Appendix E SearchOS Case Walkthroughs

The following interface snapshots are ordered by case number and then by step number. Each row contains two snapshots.

E.1 Case 1: Beijing Travel Planning

Case 1-0

Case 1-1

Case 1-2

Case 1-3

Case 1-4

Case 1-5

Case 1-6

Case 1-7

Case 1-8

E.2 Case 2: Agent Harness Engineering Survey

Case 2-0

Case 2-1

Case 2-2

Case 2-3

E.3 Case 3: GPU Cloud Provider Comparison

Case 3-0

Case 3-1

Case 3-2

Case 3-3

Case 3-4

Case 3-5