AgentFugue: Agent Scaling for Long-Horizon Tasks through Collective Reasoning

Motivation

• Unlike scaling up one agent or orchestrating agents into roles and subtasks, peer-agent scaling asks whether same-task trajectories can improve capability through cross-trajectory reuse.The introduction frames this peer setting as a distinct long-horizon reasoning problem whose gains need not come from role specialization.Source: §1 · pp. 1–2
• Peer agents can discover different reasoning paths in parallel, but isolated searches waste cross-agent progress while unrestricted communication risks noise and collapsed exploration.The proposed alternative is selective exchange that lets parallel trajectories remain distinct while developing one another’s partial progress.Source: §1 · p. 2

Framework

• AgentFugue uses an external shared reasoning hub to record concise episode notes and let agents selectively retrieve useful progress without centralized planning.The hub records what an agent established, attempted, or ruled out, while leaving local trajectories independent.Source: §1 · p. 2

Evaluation and training

• Homogeneous teams share a backbone, whereas heterogeneous teams differ in model or setup; the hub’s write and read functions are tuned by supervised fine-tuning and end-to-end reinforcement learning.The study also examines team success, individual-trajectory quality and efficiency, and conditions where communication breaks down.Source: §1 · p. 2

Results

• Across challenging benchmarks in information seeking, open-ended problem solving, and multi-step web reasoning, AgentFugue reports gains in both homogeneous and heterogeneous teams.The reported gains support peer-agent communication as a source of capability beyond stronger individual agents alone.Source: §1 · p. 2

2.1 Problem Setting

• Teams of N agents pursue the same task through distinct trajectories; AgentFugue studies homogeneous and heterogeneous teams and uses a shared hub to make intermediate discoveries reusable.The benefit depends on whether trajectories reach new task-relevant knowledge and the hub surfaces it when needed.Source: §2.1 · pp. 3–4

2.2 Shared Reasoning Hub

• At a fixed write budget, the hub summarizes a completed episode, evicts its raw content from working context, and retains it for later deep reading.Agents use visible notes to select teammate episodes and request intent-tailored synthesis of their raw content, providing broad awareness with selective depth while keeping working context bounded.Source: §2.2 · pp. 4–5

2.3 Hub Optimization

• A Qwen3.5-9B hub with separate write/read instances is trained on teacher reference notes and readouts, then aligned with GRPO while task agents remain frozen.The GRPO reward combines task success with a brevity bonus favoring outputs that lead to shorter effective search paths.Source: §2.3 · p. 5

3.1 Datasets

• Evaluation covers multi-hop evidence gathering, broad information collection, and expert multi-step reasoning, using 200-question random samples for BrowseComp and HLE and the full WideSearch benchmark.All three benchmarks follow their official judging protocols.Source: §3.1 · p. 5

3.3 Main Results

• 54.4 Avg for Qwen3.5-35B-A3B and 65.0 Avg for DeepSeek-v4-Flash make AgentFugue strongest on all three benchmarks and ahead of both multi-agent baselines.Under Qwen3.5-35B-A3B, it leads Swarm-Multi-Agent by +5.4 and Naive-Multi-Agent by +5.9; under DeepSeek-v4-Flash, it leads the corresponding baselines by +7.4 and +7.7. DeepSeek-backed AgentFugue improves BrowseComp by +15.0 (56.2→71.2), HLE by +5.5 (44.0→49.5), and WideSearch by +1.5 (72.7→74.2) over same-backbone Swarm; the Qwen-backed team also improves monotonically across all three benchmarks.Source: Table 1 · p. 6, §3.3 · pp. 6–7

3.2 Baselines

• Comparisons span single-agent ReAct, single-agent DeepResearch, and peer multi-agent systems; AgentFugue replaces centralized planning with horizontal hub exchange during exploration.Multi-agent systems share the same per-agent tool stack and interaction budget, and team predictions use the highest-confidence agent's answer.Source: §3.2 · p. 6

3.4 Scaling Behavior: Homogeneous Teams

• In homogeneous Qwen3.5-35B-A3B teams, Avg@N rises sharply before plateauing by N=5, while per-agent search and visit calls fall and per-question memory traffic rises as N grows to 8.The wide per-agent coverage band persists at the largest team size, indicating diverse trajectories despite the Avg@N plateau.Source: Figure 2 · p. 7, §3.4 · p. 7
• Pass@N, BoN@N, MV/WMV@N, and FewTool@N all improve substantially as homogeneous team size grows, while the Pass–BoN gap indicates selection headroom.BoN@N is the deployed confidence-highest selector; Pass@N reflects oracle coverage.Source: §3.4 · p. 7, Figure 3 · p. 8

3.5 Heterogeneous Teams: Stronger Models Pull Up the Group

• In heterogeneous teams, every tested aggregator improves monotonically with team size, smaller teams match the largest homogeneous Pass@N, and aggregator curves converge more tightly at the top.The tighter MV/WMV/BoN/Pass agreement leaves less headroom for smarter aggregation than in homogeneous teams.Source: Figure 3 · p. 8, §3.5 · p. 9
• Adding heterogeneous peers raises every model's performance, with weaker peers gaining double-digit margins and even Kimi gaining points; introducing a second peer initially increases browsing and memory use is several times higher than in comparable homogeneous teams.Search and visit activity contracts as the heterogeneous team grows; trajectories are not strictly monotone, as Qwen briefly stalls when GLM joins before catching up after Kimi enters.Source: Figure 4 · p. 8, §3.5 · pp. 8–9

3.6 Ablation Studies

• At N=2 on BrowseComp, every aggregator peaks at a 32K per-agent hub context budget and performs worse at smaller or larger budgets; the headline 64K trigger within a 128K context is therefore conservative.Small budgets truncate evidence before summarization, while large budgets dilute hub attention with stale content.Source: §3.6 · p. 9

Related Work

• Unlike role-specialized, orchestrated, debating, consensus, or topology-learning systems [11] [21] [5] [13] [36] [8] [4] [37], AgentFugue connects peer agents through a shared reasoning hub.The hub records what peers establish or rule out and selectively returns context to ongoing rollouts; agents do not debate or follow a fixed workflow.Source: §4 · p. 9
• Where depth- and breadth-based test-time scaling methods keep rollouts opaque during exploration [17] [29] [7] [38] [1] [35] [20] [2], AgentFugue exchanges intermediate evidence through a trained plug-in hub.Prior breadth methods merge trajectories post hoc; AgentFugue connects them while they are underway.Source: §4 · p. 9

5 Conclusion

• AgentFugue uses a shared reasoning hub to record compact notes from completed episodes and enable intent-driven access to teammate trajectories.The paper frames scaling out as a complement to strengthening a single trajectory.Source: §5 · p. 9
• Across homogeneous and heterogeneous teams, AgentFugue’s communication layer improves both team-level performance and individual trajectory quality.The results suggest peer-agent scaling can offer more than independent sampling followed by final aggregation.Source: §5 · p. 9
• A remaining challenge is making communication selective enough to preserve diversity and reliable enough to prevent misleading hypotheses from spreading.Source: §5 · p. 9

• All multi-agent systems use the same benchmark-dependent tool stack: web search and page visits for BrowseComp and WideSearch, plus Python execution and Google Scholar search for HLE.Source: A Implementation Details · p. 15
• Each agent has a 128k-token context window, and AgentFugue writes its working state to a hub episode at 64k tokens before continuing from a compressed prompt.The hub is initialized from Qwen3.5-9B and trained as described in §2.3.Source: A Implementation Details · p. 15
• Main results use each benchmark’s official judging protocol: LLM-as-a-judge for BrowseComp and HLE, and structured field-level matching for WideSearch.The exact judges and metrics are reproduced in Appendix B.Source: A Implementation Details · p. 15

B Detailed Benchmark Descriptions

• BrowseComp tests multi-hop web search, cross-document evidence aggregation, and verification; evaluation uses a fixed 200-question sample, with a shared 100-question subset for scaling and ablations.Answers are judged by the official LLM-as-a-judge protocol against gold answers.Source: B Detailed Benchmark Descriptions · p. 16
• WideSearch rewards breadth through structured field-level accuracy, while HLE [34] probes closed-book or limited-tool reasoning on a fixed 200-question sample.WideSearch uses official field-level matching; HLE uses the official LLM-as-a-judge protocol with tolerance for equivalent forms.Source: B Detailed Benchmark Descriptions · p. 16

C Detailed Baseline Descriptions

• Baseline descriptions specify each system’s underlying model, distinguishing scaffolding or training, and comparison instantiation.The appendix expands on the baselines summarized in §3.Source: C Detailed Baseline Descriptions · p. 16

C.1 LLM-based ReAct Agents

• Closed-source Claude-Opus-4.5, Kimi-K2.5, and GLM-4.7 use official APIs in the same ReAct scaffold; Qwen3.5-35B-A3B provides a like-for-like single-agent reference.All use the same web-search and page-visit tools as AgentFugue; the closed-source models retain official default decoding parameters.Source: §C.1 · p. 16

C.2 DeepResearch Agents

• WebSailor-32B [23] post-trains a 32B agent on long-horizon trajectories, while AgentFold-30B-A3B compresses reasoning history through learned folding.Their reported benchmark numbers are taken from their original papers.Source: §C.2 · p. 17
• Tongyi-DeepResearch uses iterative query planning, browsing, and summarization; OpenAI DeepResearch is evaluated using its official report’s benchmark settings.Both systems’ reported numbers are taken from their respective technical reports.Source: §C.2 · p. 17

C.3 Multi-Agent Systems

• Naive-Multi-Agent uses a meta-agent to split a task among K=2 independent ReAct subagents, then aggregate their reports without progress sharing or mid-run replanning.The meta-agent and subagents share the Qwen3.5-35B-A3B backbone.Source: §C.3 · p. 17
• Swarm-Multi-Agent lets a meta-agent create and assign reusable specialized subagents, but subagents communicate only through the meta-agent’s task reports.This allows interleaved planning and subagent invocations; the baseline uses the same backbone and per-subagent budget as the other multi-agent systems.Source: §C.3 · p. 17

D Answer Aggregation Strategies

• The strategies map agents’ extracted answers, self-reported confidence, and tool-call counts to team predictions, using the same trajectories without extra compute.They are restricted to lightweight, training-free aggregators so differences reflect how outputs are combined.Source: D Answer Aggregation Strategies · p. 18
• BoN selects the answer of the agent with the highest self-reported confidence and serves as the default aggregator in Table 1 and the main table.It uses confidence without an external reward model.Source: D Answer Aggregation Strategies · p. 18
• Majority Vote selects the most frequent answer, breaking ties by self-reported confidence.It is a baseline for measuring gains from peer agreement rather than an individual agent’s confidence calibration.Source: D Answer Aggregation Strategies · p. 18
• Weighted Majority Vote selects the answer with the largest sum of candidate agents’ self-reported confidence weights.It combines agreement with confidence, unlike BoN and unweighted Majority Vote.Source: D Answer Aggregation Strategies · p. 18
• FewTool returns the answer from the agent with the fewest tool calls, breaking ties by self-reported confidence.This data-free, efficiency-oriented heuristic tests how much accuracy is sacrificed when preferring terse trajectories.Source: D Answer Aggregation Strategies · p. 18
• Avg measures mean per-agent correctness, serving as a reference for whether an aggregator exploits the team better than a single sampled agent.Each agent’s answer is judged independently before correctness is averaged.Source: D Answer Aggregation Strategies · p. 18
• Pass@k counts a question as solved when at least one of k independently sampled agents is correct; Pass@N coincides with the Oracle.For k < N, it characterizes coverage growth from parallel rollouts and isolates sampling diversity from the selection rule.Source: D Answer Aggregation Strategies · p. 18

E Case Studies: When Shared Memory Helps and When It Misleads

• Two BrowseComp runs from the same 3-agent AgentFugue configuration show shared page memory accelerating exploration in one case and inducing confirmation bias that overrides hard constraints in another.The multi_agent_react_with_mem_v2 hub compresses evicted context into shared pages; downstream agents receive summaries and can retrieve raw content on demand.Source: E Case Studies: When Shared Memory… · p. 18

E.1 Success Case: Memory as a “Failure Map”

• Only Agent-0 reaches the gold answer, 1853; Agent-1 ends at 1885 and Agent-2 at 1848, and Agent-0 does not copy a teammate’s answer.The teammate pages Agent-0 consults explicitly say the store is still unidentified.Source: §E.1 · pp. 18–19
• Agent-0 retrieves a failure map listing candidate stores, why each was rejected, and the open question of which Eastern-Shanghai foreign-cloth store fits the clues.Reasons include candidates being too late, lacking a Yokohama Specie Bank connection, or not fitting the founder-from-Canton clue.Source: §E.1 · p. 19
• After the Hall & Holtz lead dead-ends, Agent-0 reuses the “Eastern Gate, foreign cloth, 1850s” framing to find evidence that Dafeng was created in 1853.Agent-0 checks remaining clue mismatches across steps 38–43 before committing to 1853.Source: §E.1 · p. 19
• The shared memory transports process-level state—ruled-out directions, reasons, and open sub-problems—rather than an answer, avoiding repeated dead-end searches.The passage describes this as a setting where one agent would otherwise spend its remaining budget re-litigating the same dead ends.Source: §E.1 · pp. 19–20

E.2 Failure Case: Memory-Induced Confirmation Bias

• Despite 10 memory calls, the team answered Texas Prison System Central State Farm Main Building instead of gold answer Fort Henry to an eight-constraint puzzle.The puzzle included constraints about construction, university enrollment, prison use during two wars, a preservation contributor’s father, and city population.Source: §E.2 · p. 20
• Memory contained explicit disqualifying evidence: Texas A&M enrollment and Sugar Land population missed required ranges, and the candidate was judged unlikely to fit.The step-72 return also listed failed criteria, including population growth and lack of clear evidence that Central State Farm was used as a prison.Source: §E.2 · p. 20
• The unique Gieseke clue and a retrieval goal seeking another similar candidate reframed the task, so a null result reinforced the lead rather than challenging it.The agent then treated the candidate as the “closest match” and softened failed constraints into near-satisfaction despite the recorded warnings.Source: §E.2 · pp. 20–21
• The authors propose structured candidate states such as ACTIVE, RULED_OUT, and HARD_FAIL, plus a final-answer gate that blocks candidates with recorded hard failures.They conclude that text-only page memory is insufficient for multi-constraint tasks because local uniqueness of one clue can overshadow the full constraint conjunction.Source: §E.2 · p. 21

F Key Prompts

• The page-summary prompt asks the memory manager to compress prior conversation into concise, factual continuation notes covering goals, established facts, tools, leads, failures, and open questions.It directs the manager to distinguish confirmed findings from tentative inferences and omit chit-chat and repeated content unless useful.Source: F Key Prompts · p. 21
• The page-consult prompt directs agents to extract information relevant to a research goal, preserve important prior findings, add new evidence, and distinguish confirmed from uncertain results.Its incremental template combines previous extracted results with the current page and outputs an updated focused summary.Source: F Key Prompts · p. 22
• The memory tool lets an agent retrieve raw content from up to five selected pages to recover details or verify stored explorations against a specific goal.Agents are told to inspect page summaries first and choose only pages directly relevant to the retrieval goal.Source: F Key Prompts · p. 22

Limitations

• The study uses a moderate-sized hub and limited agent backbones and configurations, leaving transfer across stronger models, model families, and hub capacities unresolved.The authors have not evaluated the full space of these models and configurations.Source: G Limitations and Broader Impact · pp. 22–23
• The evaluation covers challenging long-horizon reasoning benchmarks, not open-ended writing, sustained software engineering, or real-world interactive workflows, where framework behavior remains unvalidated.The authors believe the framework is extensible to those broader settings.Source: G Limitations and Broader Impact · p. 23
• Low-quality or overconfident notes can spread misleading conclusions, while repeated access to similar salient notes can reduce diversity and cause premature convergence.The authors identify confidence calibration, diversity-aware reading policies, and adaptive note selection as future directions.Source: G Limitations and Broader Impact · p. 23

Broader Impact

• Sharing intermediate reasoning among peer agents may benefit knowledge-intensive tasks such as scientific assistance, open-domain research, and investigative analysis.The paper frames this as a way for different exploratory paths to inform one another.Source: G Limitations and Broader Impact · p. 23
• Multi-agent systems could amplify surveillance, strategic manipulation, or deceptive content generation, and shared hubs could spread errors across a team.Suggested safeguards include access control, usage monitoring, confidence-aware hub outputs, and mechanisms that preserve diversity.Source: G Limitations and Broader Impact · p. 23

H Ablation: Hub Context-Window Size

• 32K beats the deployed 64K setting by +8.0 Pass@2, +13.5 MV@2, and +8.5 FewTool@2 on BrowseComp at N=2.Accuracy is non-monotone across budgets, peaking at 32K and degrading at both extremes: 16K cuts off useful intermediate evidence, while 96K and 128K dilute attention and create memory pressure. The 64K setting was a conservative choice intended to preserve headroom for long trajectories, not an upper bound.Source: Table 2 · p. 23, H Ablation: Hub Context-Window Size · p. 23

I LLM Usage

• LLMs serve as peer agents, the shared hub, baseline systems, and judges in experiments; paper-writing use was limited to polishing, not ideas, experiments, results, figures, tables, or analyses.Source: I LLM Usage · p. 24

1. Claims

• The authors state two core claims: a shared hub makes peer-agent count a scaling axis, and the mechanism transfers across homogeneous and heterogeneous teams.Source: §1 · p. 25

2. Limitations

• The paper identifies limited backbone and hub-capacity coverage, restriction to long-horizon QA-style benchmarks, and collective-reasoning failure modes including low-quality notes and premature consensus.Source: §2 · p. 25

3. Theory assumptions and proofs

• The paper is empirical and contains no formal theorems or proofs.Source: §3 · p. 26

4. Experimental result reproducibility

• The paper reports that it specifies the architecture and hub-optimization recipe, tool stack and settings, benchmark splits, judging protocols, and baseline configurations for reproducing the comparisons.Source: §4 · p. 26

5. Open access to data and code

• Code and trained hub checkpoints were withheld to preserve anonymity; benchmarks are public, algorithmic and configuration details are provided, and code release is planned for the camera-ready version.Source: §5 · p. 27

6. Experimental setting/details

• Experimental details include benchmark setup, baseline groups, team-size sweeps, a 150-round interaction budget, a 128k agent context, a 64k hub-write trigger, and hub initialization.Source: §6 · p. 27

7. Experiment statistical significance

• The paper reports point-estimate accuracy rather than confidence intervals, following benchmark convention [34], and plans bootstrap intervals for per-question outcomes in the camera-ready version.Team-size and aggregation-rule sweeps characterize variability indirectly.Source: §7 · p. 28

8. Experiments compute resources

• Appendix A documents per-agent context windows, interaction budgets, and hub triggers; training hardware and inference wall-clock costs are deferred to the camera-ready version.Source: §8 · p. 28

9. Code of ethics

• The authors report reviewing the NeurIPS Code of Ethics and using public benchmarks and pretrained models without personal or sensitive data or human-subjects experiments.They state that anonymity is preserved throughout the submission.Source: §9 · p. 28

10. Broader impacts

• The paper discusses benefits of more capable assistants and risks from coordinated misuse or propagated erroneous conclusions, with access control, monitoring, and diversity-preserving hub policies as mitigations.These impacts are discussed in Section G (“Broader Impact” paragraph).Source: §10 · p. 29

11. Safeguards

• The submission releases no high-risk asset or new public dataset, model checkpoint, or code package; the trained hub checkpoint is not included, though it may be released later.Source: §11 · p. 29, §13 · p. 30

12. Licenses for existing assets

• The paper cites third-party benchmarks and pretrained backbones at first use, uses them under their original licenses and terms, and redistributes no assets.Source: §12 · p. 30

14. Crowdsourcing and research with human subjects

• The paper uses existing public benchmarks with automated evaluation and involves no crowdsourcing or human-subjects research, so IRB review does not apply.Source: §14 · pp. 30–31

16. Declaration of LLM usage

• LLMs are central to the method and experiments: the shared hub is trained by supervised fine-tuning and reinforcement learning, and the paper documents agent and baseline models and judge protocols.The authors state that LLM use in paper writing was limited to language polishing.Source: §16 · p. 31