Source-linked AI summary
SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
Xiaochong Lan, Quan Chen, Kun Tao, Xinyu Tang, Tianshu Wang, Qianggang Cao, Xinyu Kong, Zujie Wen, Zhiqiang Zhang, Jun Zhou
TL;DR
Long-horizon agents face finite context windows and scarce training data for deciding what, when, and how to delegate. SearchSwarm uses a harness to elicit and distill delegation behavior into model weights, achieving state-of-the-art results among comparable-scale models across four benchmarks.
Problem
Long-horizon agents face finite context windows, while training data for delegation intelligence remains scarce and its synthesis and training are largely unexplored.
Method
A harness guides task decomposition, comprehensive subagent briefing, and result integration, whose trajectories are distilled into supervised fine-tuning data.
Results
SearchSwarm achieves state-of-the-art performance among comparable-scale models across BrowseComp, BrowseComp-ZH, GAIA, and xbench-DeepSearch, scoring 68.1 on BrowseComp.
Takeaways & Limitations
The trained delegation intelligence generalizes to single-agent settings and open-ended research tasks within the demonstrated deep-research scope.
Abstract
from arXiv · showhide
Large language models are increasingly expected to handle complex, long-horizon real-world tasks whose context demands can grow without bound, yet model context windows remain inherently finite. Recent work explores a paradigm where a main agent decomposes tasks and dispatches subtasks to subagents, which execute and return only summarized results, conserving the main agent's context budget. However, performing this well requires delegation intelligence: the ability to decompose complex tasks, determine when and what to delegate, and integrate returned results into the ongoing workflow. Training data for this capability is scarce in naturally occurring text, and to our knowledge, how to synthesize such data and train models to acquire this capability remains largely unexplored in the open-source community. To bridge this gap, we present a preliminary exploration targeting deep research, a representative long-horizon agent task. Specifically, we design a harness that guides the model toward high-quality task decomposition and delegation, while constraining subagents to return results properly to support the main agent's workflow. The harness-guided trajectories naturally encode correct delegation decisions, which we use as supervised fine-tuning data to internalize delegation intelligence into model weights. Our resulting model, SearchSwarm-30B-A3B, achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, the best results among all models of comparable scale. We will release our harness, model weights, and training data to facilitate future research.
1 Introduction
SearchSwarm addresses the finite-context challenge of long-horizon agent tasks through delegation intelligence: decomposing work, selecting subtasks, and integrating summarized subagent results. It uses a harness to synthesize supervised fine-tuning data that internalizes these behaviors, achieving strong benchmark results and releasing the resulting resources.
- Motivation: Long-horizon agent tasks can have unbounded context demands despite inherently finite model context windows, motivating selective information retention or condensation.The introduction identifies context management as a fundamental requirement for complex real-world tasks.
- Delegation intelligence: Delegation intelligence enables a main agent to decompose complex tasks, decide when and what to delegate, and integrate subagent results into its workflow.Training data for this capability is scarce because natural corpora rarely show explicit multi-agent coordination.
- Harness: The proposed harness elicits high-quality delegation during deep research by dispatching bounded subtasks to parallel subagents while preserving the main agent’s independent understanding of research progress.The harness uses a call_sub_agent tool and requires subagent briefings to include task rationale and relevance.
- Training: Filtered harness-guided trajectories provide supervised fine-tuning data encoding correct decisions about decomposition, subtask scope, and subagent briefing.Fine-tuning internalizes these decision patterns into model weights, enabling models initially lacking delegation intelligence to exhibit the behavior.
- Results and release: 68.1 on BrowseComp, 73.3 on BrowseComp-ZH, 82.5 on GAIA, and 80.8 on xbench-DeepSearch are achieved by SearchSwarm, the best results among models of comparable scale.The paper identifies SearchSwarm-30B-A3B as the resulting model and reports these benchmark results as a contribution.
- Results and release: The authors open-source the harness, model weights, and training data to facilitate research on delegation intelligence and multi-agent coordination.The release is intended to support future research.
2 Method
SearchSwarm organizes deep research around a main agent that delegates bounded subtasks to independent same-model subagents and integrates their condensed, citation-grounded reports. A harness guides delegation, briefing, judgment retention, and reporting, while successful trajectories are collected for supervised fine-tuning.
- 2.1 Formalization: The main agent plans and delegates bounded subtasks, while independent subagents execute them and return condensed reports for integration.Subagents operate in fresh contexts and the main agent sees only their final reports, not intermediate execution steps.
- 2.2 Harness Design: The harness encourages delegation of token-expensive information gathering, requires comprehensive briefs, preserves the main agent’s directional judgment, and enforces citation-grounded reports.The main agent independently decides which hypotheses to pursue, when to terminate, and how to adjudicate conflicting reports.
- 2.1 Formalization: Delegation manages context by passing subagents only a brief and reintroducing a compressed report rather than the full interaction history.The subagents are independent invocations of the same model, not separate additional models.
- Data Collection: Training data consists of complete harness-guided deep-research trajectories sourced from the RedSearcher and OpenSeeker datasets, retaining forced-answer trajectories for test-time behavior.The data collection targets both effective delegation and execution of delegated tasks.
- Filtering and Training: Filtering retains correct main-agent trajectories, keeps corresponding subagent trajectories, downsamples overly short subagent runs, and removes repeated calls, fabricated citations, and tool misuse.The training objective fine-tunes the base model by next-token prediction with environment masking over alternating model outputs and environment returns.
3 Experiments
Experiments evaluate SearchSwarm on four long-horizon research benchmarks and show state-of-the-art performance among 30B-A3B models, with competitiveness against larger systems. Ablations and controlled comparisons indicate that delegation behavior depends on harness-guided training and generalizes beyond tool-enabled settings.
- Main results: SearchSwarm achieves state-of-the-art performance among 30B-A3B models across BrowseComp, BrowseComp-ZH, GAIA, and xbench-DeepSearch-2505.It scores 68.1, 73.3, and 82.5 on the first three benchmarks, and 80.8 on xbench-DeepSearch-2505.
- Main results: 24.7 points on BrowseComp separate SearchSwarm from the base model without context management, rising from 43.4 to 68.1.The comparison uses BrowseComp scores for SearchSwarm and its base model without context management.
- Main results: 82.5 on GAIA lets SearchSwarm surpass GPT-5 (76.4) and Seed-2.0-Pro (78.6), while trailing only Step-3.5-Flash (84.5).On BrowseComp, it also matches DeepSeek V3.2 (671B-A37B, 67.6) and exceeds GPT-5.2-Thinking (65.8).
- Harness ablation: 57.7 on a 200-question BrowseComp subset is achieved by the full harness, versus 50.0 with only the delegation-tool schema and 47.7 for the original framework.Providing the delegation tool alone yields a +2.3 improvement, while the full harness adds design principles for delegation, briefing, and citations.
- Training-data comparison: 66.5 on BrowseComp and 64.0 on BrowseComp-ZH are achieved by Qwen3-30B-A3B-Thinking-2507 fine-tuned on the same delegation data.This controlled experiment isolates the contribution of the training data under the same experimental setup as the main experiments.
- Generalization: 52.0 and 53.3 on BrowseComp and BrowseComp-ZH exceed Tongyi DeepResearch’s 43.5 and 46.5 without the subagent tool.Both models use an identical single-agent configuration with one 128K-token context, no context management, and the call_sub_agent tool disabled.
- Delegation behavior: Over 70% of main-agent tool calls on BrowseComp and BrowseComp-ZH invoke call_sub_agent, compared with 43–51% on GAIA and xbench.Direct main-agent tool use is verification-oriented, while subagents search predominantly at rates of 46.5–76.6%.
4 Related Work
Prior work frames delegation as a way to overcome individual cognitive limits, and recent LLM-agent research applies this principle to finite context windows through coordinated subagents. Related work also characterizes agents as tool-using systems that interact with environments, while search provides access to information beyond fixed model parameters.
- Delegation: Human delegation manages complexity by assigning subtasks and integrating results, requiring judgments about when and how to delegate.The passage distinguishes effective delegation from mere task forwarding.
- Delegation: LLM agents use delegation to address context limits by offloading subtasks to independent instances and receiving condensed results.Recent architectures include coordinators that dispatch focused subagents in parallel and synthesize their reports.
- Agentic LLMs: Modern LLM agents extend beyond single-turn question answering by using tools, interacting with environments, adapting to feedback, and completing multi-turn tasks.The passage situates this development across several contemporary model families.
- Search: Because model parameters are a lossy, fixed representation of world knowledge, search provides access to real-time and long-tail information relevant to decisions.The passage links information value to improved decision-making.
5 Conclusion
SearchSwarm presents a preliminary approach to training delegation intelligence for long-horizon agent tasks, demonstrated on deep research. Its harness improves inference-time research performance and generates supervised fine-tuning data that internalizes delegation behavior into model weights.
- SearchSwarm targets delegation intelligence for long-horizon agent tasks and demonstrates its effectiveness on deep research.
- The harness guides task decomposition, comprehensive subagent briefing, and citation-grounded result integration.
- The harness improves deep research performance at inference time.
- Harness-generated supervised fine-tuning data internalizes delegation behavior into model weights.
- SearchSwarm-30B-A3B achieves state-of-the-art performance.
A Behavioral Analysis · B Full Prompts
Behavioral distributions distinguish successful from unsuccessful research trajectories, with incorrect answers associated with broader and sometimes much longer agent interaction patterns. The full prompts enforce evidence-grounded answers, explicit candidate comparison, citation integrity, uncertainty handling, and structured tool use for main and subagents.
- A Behavioral Analysis: Correct answers concentrate in moderate subagent-call ranges, while incorrect answers show flatter distributions extending to much higher call counts.Peaks occur at 2–3 calls on GAIA and xbench, and 3–5 on BrowseComp and BrowseComp-ZH.
- A Behavioral Analysis: Incorrect answers use broader main-agent turn distributions, including a secondary BrowseComp peak around 20–30 turns.Correct answers peak at 3–5 turns on BrowseComp and BrowseComp-ZH, and 2–4 on GAIA and xbench.
- A Behavioral Analysis: Subagent turn distributions have similar peak positions across benchmarks, but BrowseComp exhibits a pronounced peak at the 50-turn limit.Correctly answered questions otherwise show a more uniform subagent-turn distribution.
- B Full Prompts: The main-agent prompt requires explanations that identify every entity and constraint, justify each inference, and attach inline citations to supporting evidence.The explanation is designed for questioners with zero background so they can verify answers at low cost.
- B Full Prompts: The grounding instructions require every claim to be supportable from retrieved search or visit passages, with unsupported points flagged rather than fabricated.The main agent must continue researching until the grounding bar is met and explicitly resolve ambiguities with retrieved citations.
- B Full Prompts: When multiple candidates remain, the prompt requires side-by-side evidence for and against each candidate and explicit reasons for selection and rejection.This comparison requirement applies at delivery when more than one candidate is still viable.
- B Full Prompts: Citation instructions reject adjacency, nearby support, non-trivial inference, fabricated URLs, and suggestive snippets as sufficient evidence.Search-snippet citations must be marked “(search snippet)” only when the snippet directly supports the claim, and references must be listed in order.
- B Full Prompts: The main agent can dispatch explicit parallel research assignments, whereas subagents receive task briefs and lack access to call_sub_agent.Both agents use search, visit, PythonInterpreter, and google_scholar tools, with calls serialized as JSON inside tool_call tags.
C Case Study
A representative harness trajectory answers a difficult multi-constraint question correctly by identifying the Coomera Connector (M9) motorway in Queensland after two delegation rounds, exhibiting all four design principles from Section 2.2.
- Representative trajectory: The model correctly identifies the Coomera Connector (M9) motorway in Queensland after two rounds of delegation on a hard, multi-constraint question.The trajectory demonstrates all four design principles described in Section 2.2.
User question
The session demonstrates delegation-first research: the main agent decomposes the clues, dispatches parallel sub-agents across three investigative angles, and synthesizes their reports. All three sub-agents converge on the Coomera Connector, while the main agent remains responsible for orchestration and synthesis.
- Delegation flow: The session proceeds in two delegation rounds with the main agent acting as orchestrator throughout.Delegation crosses from sub-agent briefs and reports back into the main agent’s reasoning.
- Delegation flow: The main agent decomposes seven clues into three clusters and dispatches one parallel sub-agent per cluster.The clusters concern the motorway’s northern section, the national rail upgrade, and the project’s former corridor name.
- Delegation flow: All three sub-agents independently converge on the Coomera Connector, after which the main agent synthesizes their reports.The supplied passage indicates that the main agent also detects a subsequent issue, but that reasoning is truncated.
- Delegation flow: After parsing the clues, the main agent fans out instead of searching alone, issuing three briefs to investigate multiple angles in parallel.Its initial reasoning hypothesizes that the rail upgrade is likely Inland Rail and reserves the main context for synthesis rather than raw retrieval.
Main-agent reasoning (first step) … 1. INLink JV — Parkes to Narromine Section of Inland Rail
The main agent decomposes a multi-clue infrastructure identification task, dispatches targeted verification briefs, and critically evaluates convergent sub-agent answers. The investigation identifies the Coomera Connector while verifying Fulton Hogan’s leadership of the INLink JV for Inland Rail’s Parkes–Narromine section.
- Main-agent reasoning (first step): The main agent decomposes the question into six explicit clues covering funding, geography, project status, consortium history, and corridor origins.The clues target a major Eastern Australian motorway and its connection to a national rail upgrade.
- Brief dispatched to a sub-agent: The verification brief gives the sub-agent established context and asks whether Fulton Hogan led the INLink JV and connected the two projects.The brief is designed to prevent duplicated investigation and focus the sub-agent on the unresolved relationship.
- Brief dispatched to a sub-agent: The sub-agent is explicitly tasked with checking leadership of the Inland Rail JV, leadership of Coomera Connector Stage 1 Central, and the FHHMJV composition.A separate instruction requests detailed findings on both joint ventures’ leadership structures.
- Main-agent reasoning (core judgment): The main agent rejects unverified convergence, distinguishes the FHHMJV from the INLink JV, and identifies the Premier funding clue as requiring further verification.Only Fulton Hogan is common to both consortiums, rather than the consortiums being identical.
- Main-agent reasoning (core judgment): All three sub-agents converge on the Coomera Connector (M9) in Queensland as the answer.The reported constraints include equal federal–Queensland funding and the opening of Stage 1 North.
- Main-agent reasoning (core judgment): The main agent’s final explanation uses citation-grounded sub-agent reporting to resolve seven constraints and compare alternative candidates with end-to-end traceability.Inline citations allow each substantive claim to be checked against retrieved evidence.
- 1. INLink JV — Parkes to Narromine Section of Inland Rail: INLink was a joint venture between BMD Constructions and Fulton Hogan delivering the $310 million Parkes to Narromine section of Inland Rail.Fulton Hogan’s project page identifies the INLink JV and names Gerard O’Connor from Fulton Hogan as Project Director.
- 1. INLink JV — Parkes to Narromine Section of Inland Rail: Fulton Hogan led the INLink JV for the Parkes to Narromine Inland Rail project.The cited project description states that the JV was headed by Gerard O’Connor, Project Director from Fulton Hogan.
2. FHHMJV — Coomera Connector Stage 1 Central
FHHMJV is the Fulton Hogan-led consortium of Fulton Hogan, JF Hull Holdings Pty Ltd, and McIlwain that won both Coomera Connector Stage 1 Central contracts. Sources differ on when its early works contract was awarded, citing January or May 2023.
- JV Composition: FHHMJV comprises Fulton Hogan, JF Hull Holdings Pty Ltd, and McIlwain, with Fulton Hogan as lead.The acronym expands to Fulton Hogan JF Hull Holdings Pty Ltd McIlwain Joint Venture.
- Contract Awards: The consortium won both the early works and main construction contracts for Coomera Connector Stage 1 Central.The consortium was awarded both contracts for the package.
- Early Works Contract: January 2023 and May 2023 are the reported award dates for the early works contract, depending on the source.Wikipedia and AARoads Wiki report January 2023, while Infrastructure Pipeline reports May 2023.
3. Summary Comparison · Key Finding
The comparison contrasts the INLink JV and FHHMJV across partners, contract value, timing, and project status. Fulton Hogan is the confirmed lead company in both joint ventures.
- 3. Summary Comparison: The comparison covers INLink JV (Inland Rail P2N) and FHHMJV (Coomera Connector Stage 1 Central).
- Key Finding: Fulton Hogan leads both projects, partnering with BMD Constructions in INLink JV and JF Hull Holdings Pty Ltd and McIlwain in FHHMJV.
- 3. Summary Comparison: ~$310 million is the contract value for INLink JV, compared with ~$1.5 billion for the FHHMJV Stage 1 Central portion.
- 3. Summary Comparison: October 2018 marks the INLink JV main contract award, while the FHHMJV main contract award occurred in September 2024.
- 3. Summary Comparison: December 13, 2018 is the INLink JV groundbreaking date, whereas FHHMJV main works began in September 2024.
- 3. Summary Comparison: INLink JV is completed and operational since late 2020, while FHHMJV remains under construction.
Final explanation delivered to the user
The answer identifies the Coomera Connector (M9) in Queensland as the motorway satisfying the question’s constraints. It supports this identification through the project’s partial-completion status and by excluding several alternative motorway candidates.
- Identification: The project is the Coomera Connector (M9), a motorway on the Gold Coast in Queensland.Queensland is identified as an Eastern Australian state, and the explanation frames the project as satisfying the specified constraints.
- Constraint verification: The Coomera Connector is a partially completed 45-kilometre motorway connecting Loganholme to Nerang.As of late 2025, only the 4-km northern section was open, while the remaining sections were still under construction.
- Alternative elimination: The explanation rules out the M12 Motorway, Coffs Harbour Bypass, West Gate Tunnel, and M1 Pacific Motorway extension as mismatches.The alternatives differ from the requested description in construction timing, opening dates, corridor naming, form, cost, or section-opening status.