SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research

Abstract: LLMs 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.

• The paper introduces a harness and supervised fine-tuning recipe that teaches a 30B-A3B model to decompose research tasks, brief subagents, verify reports, and synthesize answers, achieving 68.1 on BrowseComp, 73.3 on BrowseComp-ZH, 82.5 on GAIA, and 80.8 on xbench-DS.
• SearchSwarm improves BrowseComp by 24.7 absolute points over its base model, while ablations show that comprehensive delegation prompts contribute more than merely adding a subagent tool and that prompting alone does not induce delegation.
• The learned investigative behaviors generalize beyond delegated execution, improving open-ended research performance by 14.2 points on average and raising single-agent BrowseComp results, although compute fairness, recursive delegation, and human-verified evaluation remain open issues.

Overview and motivation

SearchSwarm addresses a specific capability gap in agentic LLMs: delegation intelligence, defined as the main agent's ability to decompose complex tasks, decide when and what to delegate to subagents, brief those subagents effectively, and integrate their returned reports into an ongoing workflow. The authors motivate this as a form of active context management. Long-horizon tasks such as deep research generate context demands that grow without bound, while context windows remain finite; existing remedies (history truncation, fixed-rule summarization, retaining only recent tool outputs) are characterized as passive because they react only after the budget is exhausted or discard information indiscriminately.

The paper's central observation is that training data exhibiting explicit multi-agent coordination is scarce in naturally occurring text, and that the open-source community lacks a complete recipe—harness design, data synthesis, and model training—for acquiring delegation behavior. SearchSwarm is presented as a preliminary exploration of such a recipe, instantiated on deep research, with all components (harness, weights, training data) slated for open release.

Method

Formulation

The task is modeled as a ReAct-style trajectory $$H_T = (q, (\tau_0,a_0,o_0),\ldots,(\tau_T,a_T,o_T), y)$$ over a tool-equipped environment. Delegation occurs when $a_t = \texttt{call\_sub\_agent}(b)$: the brief $b$ triggers an independent sub-trajectory conditioned solely on $b$, producing a report $r$ that becomes the main agent's observation. Subagents have no visibility into the main agent's history, and the main agent sees only the final report.

A notable framing choice: although the architecture involves subagents, it uses a single model invoked in fresh contexts rather than separate models. The authors therefore characterize the method as single-agent context management in which model-generated briefs and reports act as content-aware compression replacing fixed-rule truncation. This framing is used to justify comparing against other context-management approaches "on equal footing"—a claim worth scrutinizing, since token cost per query is not equivalent across these mechanisms (parallel subagents consume substantially more total compute than truncation).

Harness design

The harness consists of tools (search, visit, google_scholar, python, plus call_sub_agent) and system prompts built on four principles:

• Encouraging delegation: multi-step information gathering is token-expensive but cognitively shallow; the main agent should offload it and reserve its context for planning, verification, and synthesis.
• Comprehensive briefing: each brief must include why the subtask matters, what is established, what remains uncertain, and which directions were ruled out—treating the subagent as a new collaborator joining mid-investigation.
• Main agent retains core judgment: subagents gather evidence; directional decisions (hypothesis selection, termination, adjudication of conflicts) stay with the main agent.
• Citation-grounded reporting: every important subagent conclusion carries inline citations to source URLs, enabling verification despite the main agent's inability to observe subagent execution.

Delegation is limited to one level: subagents lack call_sub_agent.

Supervised fine-tuning

Trajectories are collected from queries sourced from RedSearcher and OpenSeeker under two configurations: (1) a single model playing both roles, retaining both trajectories; (2) a stronger main agent paired with a weaker subagent, retaining only main-agent trajectories—the rationale being that unreliable subagent results force more deliberate decomposition and verification. Context limits are 128K (main) and 64K (sub). Forced-answer trajectories near the context limit are deliberately retained so the model learns to respond well when this mechanism triggers at test time.

Filtering keeps only main-agent trajectories with correct final answers, removes degenerate behaviors (repeated identical tool calls, hallucinated citations, web access via the Python interpreter), and downsamples overly short subagent trajectories. Training is next-token prediction over model outputs with environment returns masked, applied uniformly to both roles.

Main results

SearchSwarm-30B-A3B, fine-tuned from Tongyi DeepResearch-30B-A3B, achieves state-of-the-art results among 30B-A3B models on all four benchmarks:

Model	Size	BrowseComp	BrowseComp-ZH	GAIA	xbench-DS
MiroThinker-1.7-mini	30B-A3B	67.9*	72.3*	80.3*	—
LongSeeker	30B-A3B	61.5*	62.5*	77.7*	78.0*
Tongyi DeepResearch (base)	30B-A3B	43.4	46.7	70.9	75.0
SearchSwarm	30B-A3B	68.1*	73.3*	82.5*	80.8*

(* indicates results with context management.)

The headline claims are strong: relative to the base model without context management, fine-tuning yields a 24.7-point absolute improvement on BrowseComp, which the authors attribute directly to internalized delegation intelligence. Against much larger systems, SearchSwarm matches DeepSeek V3.2 (671B-A37B, 67.6) and exceeds GPT-5.2-Thinking (65.8) on BrowseComp, and surpasses GPT-5 (76.4) and Seed-2.0-Pro (78.6) on GAIA, trailing only Step-3.5-Flash (84.5).

An instructive negative control: applying the harness to the base model without fine-tuning ("Tongyi DR Swarm") produces no behavioral change—the model never invokes call_sub_agent and performs identically to the base model. This establishes that delegation behavior does not emerge from prompting alone and requires weight-level training, supporting the paper's premise that this capability must be explicitly taught.

Ablations and generalization

Harness effectiveness. On a 200-question BrowseComp subset using DeepSeek V3.2 as the agent: the original Tongyi framework scores 47.7; adding call_sub_agent with only its schema described yields 50.0 (+2.3); the full harness yields 57.7 (+10.0 over base). The design principles—not merely the tool's existence—drive most of the gain, and invocation frequency rises markedly under the full harness.

Training-data portability. Fine-tuning Qwen3-30B-A3B-Thinking-2507, a model never optimized for deep search, on the same data yields 66.5 on the 200-question BrowseComp subset and 64.0 on BrowseComp-ZH, exceeding RedSearcher's and LongSeeker's reported best configurations. This isolates the contribution of the synthesized data itself. One caveat: RedSearcher reports on the full 1266-question benchmark, so that comparison crosses different evaluation subsets, though the margin makes the ordering likely robust.

Single-agent generalization. With call_sub_agent disabled and no context management, SearchSwarm scores 52.0/53.3 versus the base model's 43.5/46.5 on the BrowseComp subset/BrowseComp-ZH—despite no single-agent trajectories in training. This suggests the encoded investigative patterns (decomposition, systematic sub-question resolution, progress tracking) transfer beyond the delegation setting.

Open-ended generalization. On four long-form benchmarks (ScholarQA-v2, HealthBench, ResearchQA, DeepResearchBench), SearchSwarm averages 64.2 versus the base model's 50.0 (+14.2), with ScholarQA-v2 improving +32.7. It trails Dr.Tulu (65.6) among open-source systems and essentially matches OpenAI DeepResearch (64.9)—again trained exclusively on short-answer queries, indicating the structured investigative process transfers to synthesis-heavy tasks.

Behavior analysis

Tool-use distributions confirm the intended division of labor: call_sub_agent accounts for over 70% of main-agent invocations on BrowseComp and BrowseComp-ZH (43–51% on GAIA and xbench); direct main-agent tool use is verification-oriented, with visit disproportionately dominant (e.g., 26.4% visit vs. 11.1% search on GAIA), consistent with following citation URLs from subagent reports. Subagents show the inverse pattern, dominated by search (46.5–76.6%). Python usage concentrates in the main agent on computation-heavy benchmarks (GAIA, xbench). Appendix statistics further show correctly answered questions cluster at moderate subagent-call counts (2–5) and turn counts, while failures spread toward far higher counts—a correlation, not necessarily a causal diagnostic of failure modes.

Limitations and open questions

Several constraints bound the strength of the conclusions. First, the evaluation relies on LLM-as-judge (DeepSeek-V4-Flash) with manual spot-checking rather than exhaustive human verification, and several comparisons use 200-question subsets due to resource constraints. Second, the compute-fairness question raised by the single-agent framing remains unaddressed: parallel subagents multiply inference cost, and no wall-clock or token-budget accounting against truncation-based baselines is provided. Third, delegation is capped at one level; whether recursive delegation scales further is untested. Fourth, the approach depends on answer-correctness filtering, which ties training-signal quality to benchmark-verifiable queries and may limit applicability to domains lacking ground truth. Finally, the work is explicitly framed as preliminary: reinforcement learning over delegation policies (as in Agent Swarm) is not explored, leaving open whether SFT alone captures the full ceiling of delegation intelligence, and how the learned policy behaves when subagent reliability distributions shift at deployment time.

Conclusion

SearchSwarm provides a complete, openly released recipe—harness, filtered SFT data, and weights—for training delegation intelligence into a 30B-scale model, achieving the best reported results among comparable-scale models on four short-answer deep-research benchmarks (68.1 BrowseComp, 73.3 BrowseComp-ZH, 82.5 GAIA, 80.8 xbench-DeepSearch) while remaining competitive with frontier systems an order of magnitude larger. The ablations substantiate two non-obvious findings: harness design principles contribute most of the inference-time gain over merely exposing a delegation tool, and the resulting capabilities generalize to single-agent execution and open-ended synthesis despite absent training coverage. The principal unresolved issues concern compute-matched comparisons, deeper delegation hierarchies, and post-SFT optimization of delegation policy.