PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yuyang Hu, Hongjin Qian, Shuting Wang, Jiongnan Liu, Tong Zhao, Xiaoxi Li, Zheng Liu, Zhicheng Dou

arXiv:2605.24486v1cs.AIcs.CL

TL;DR

The paper asks whether peer agents exploring the same long-horizon task can gain capability through shared reasoning rather than isolated search. AgentFugue uses a shared reasoning hub for selective cross-agent communication and improves performance across the evaluated settings.

  • Problem

    Evidence is limited on whether multiple peer agents exploring the same long-horizon task can provide capability gains beyond scaling a single trajectory or adding compute.

  • Method

    AgentFugue uses a shared reasoning hub that records compact episode notes and supports intent-driven retrieval of teammates’ progress without centralized planning.

  • Results

    Across homogeneous and heterogeneous teams and multiple long-horizon benchmarks, AgentFugue improves performance; under DeepSeek-v4-Flash it reaches 65.0 Avg, +7.4/+7.7 over corresponding baselines.

  • Takeaways & Limitations

    The results suggest peer-agent scaling with shared intermediate reasoning can provide capability gains beyond independent sampling followed by final aggregation.

  • Takeaways & Limitations

    Low-quality or overconfident notes may spread misleading conclusions, while repeated reading of salient notes may reduce trajectory diversity and cause premature convergence.

Abstract

from arXiv · show

Recent progress on long-horizon agentic tasks has been driven largely by scaling up individual agents through stronger models, better tools, and more effective scaffolding. In contrast, much less is understood about scaling out: whether multiple peer agents, all targeting the same task, can become an additional source of capability without relying on explicit role specialization or workflow orchestration. We study this question and propose AgentFugue, a collective reasoning framework built around a shared reasoning hub. As peer agents explore the same task in parallel, the hub records concise notes on what each agent has established, attempted, or ruled out, and enables each agent to selectively access what other agents have discovered in a form useful for its current search. This design turns otherwise isolated trajectories into a connected ecology of reusable intermediate reasoning without requiring centralized planning. We instantiate the hub as a plug-in communication layer, trained with supervised fine-tuning and end-to-end reinforcement learning. Across the challenging long-horizon settings we study, AgentFugue improves over strong baselines. Our results suggest that collective reasoning can turn scaling out peer agent systems into a distinct source of capability gains, rather than merely a way of spending more compute.

1 Introduction

The introduction frames scaling out peer agents on the same long-horizon task as a distinct capability source beyond scaling individual agents or adding compute. It presents AgentFugue, a shared reasoning hub that enables selective cross-trajectory reuse without centralized planning, and reports gains across homogeneous and heterogeneous teams.

  • Motivation: Scaling individual agents through stronger models, better tools, and improved scaffolding has expanded their long-horizon capabilities, but scaling out remains less understood.The paper contrasts scaling up with increasing the number or diversity of peer agents working on the same task.
  • Motivation: Unlike multi-agent approaches centered on role assignment, task decomposition, or explicit workflows, AgentFugue studies parallel peer exploration of the same task.The introduction positions peer-agent scaling as capability arising from cross-trajectory reuse rather than role specialization.
  • AgentFugue: AgentFugue uses a shared reasoning hub that records compact notes about agents’ established, attempted, or ruled-out progress and lets others selectively retrieve useful information.The hub is an external communication layer rather than a centralized planner, turning parallel trajectories into a shared reasoning ecology.
  • AgentFugue: The framework supports homogeneous teams, where gains must come from interaction among similar agents, and heterogeneous teams, where differing reasoning biases may complement one another.These settings isolate interaction effects from built-in role differences while also testing complementarity across models or setups.
  • Results: AgentFugue gains across information seeking, open-ended problem solving, and multi-step web reasoning benchmarks in both homogeneous and heterogeneous teams.The reported results support peer-agent communication as a capability source beyond stronger individual agents alone.

2 Method

AgentFugue models multiple peer agents exploring the same long-horizon task as distinct trajectories whose intermediate discoveries can be connected through a shared reasoning hub. The hub compresses completed episodes into reusable notes and supports intent-driven access to relevant teammate work without centrally orchestrating search.

  • Task, team, and trajectories: A team of N agents targets the same task, with each agent producing a distinct exploratory trajectory and discovered subspace.The framework includes homogeneous teams with shared models and configurations, and heterogeneous teams that differ in backbone or prompting.
  • Shared reasoning hub: The shared reasoning hub connects trajectories by compressing completed reasoning segments into reusable notes that expose established evidence, attempts, and ruled-out branches.Writing an episode both makes partial progress visible to teammates and evicts raw content from the working context.
  • Episode writing and context eviction: The working context combines an agent’s own episode notes, visible teammate notes, and its current unfinished interaction, keeping context bounded as reasoning effort grows.This provides a structured view of collective progress while allowing local reasoning to continue.
  • Intent-driven reading: Agents retain coarse awareness through episode notes and selectively request deeper synthesis of referenced teammate episodes using an intent-driven read.The resulting readout is tailored to the requesting agent’s current need rather than passively exposing all teammate content.
  • Distinction from nearby paradigms: AgentFugue differs from debate, group chat, best-of-N sampling, and single-agent memory because agents asynchronously reuse intermediate progress before task completion.The mechanism is designed for cross-agent reuse without synchronized turn-taking or a shared conversational context.

3 Experiments

AgentFugue is evaluated on three complementary long-horizon benchmarks and outperforms multi-agent baselines across all of them. Scaling experiments show benefits from larger homogeneous and heterogeneous teams, shared-hub coordination, aggregation, and an intermediate context budget.

  • Benchmarks: The evaluation spans BrowseComp’s deep web search, WideSearch’s breadth-oriented evidence collection, and HLE’s expert-authored multi-domain reasoning.These benchmarks stress retrieval, breadth, and reasoning capabilities respectively.
  • Main results: AgentFugue leads every benchmark and beats multi-agent baselines by +5.4/+5.9 under Qwen3.5-35B-A3B and +7.4/+7.7 under DeepSeek-v4-Flash.Its average scores are 54.4 and 65.0, respectively, with gains attributed to shared-hub coordination.
  • Main results: +15.0 on BrowseComp, +5.5 on HLE, and +1.5 on WideSearch versus the same-backbone Swarm baseline show gains across retrieval, reasoning, and breadth.For DeepSeek, scores move 56.2→71.2, 44.0→49.5, and 72.7→74.2.
  • Aggregation: Every tested aggregator improves substantially with team size, so scaling does not depend on the selector.The aggregators are Pass@N, BoN@N, MV/WMV@N, and FewTool@N.
  • Heterogeneous scaling: Heterogeneous teams benefit every backbone, especially weaker peers, while complementary models initially increase exploration and generate several-times-higher hub memory traffic.Aggregator-side scaling remains monotonic, with faster gains per added peer than in homogeneous teams.
  • Hub budget: At N=2, every aggregator follows an inverted-U over hub budgets, peaking at 32K and degrading at both smaller and larger context windows.Small budgets truncate evidence, whereas large budgets dilute attention with stale content.

4 Related Work

Prior multi-agent LLM systems typically rely on role specialization, debate or consensus, and orchestrated or learned interaction structures. Test-time scaling expands either trajectory depth or breadth, whereas AgentFugue uses peer agents that exchange intermediate evidence through a shared reasoning hub during exploration.

  • Multi-agent LLM collaboration: Prior multi-agent LLM collaboration commonly specializes agent roles and orchestrates their interaction through typed frameworks, debate or consensus, or learned topologies.AgentFugue instead uses peer agents, avoids debate and fixed workflows, and shares a common reasoning hub.
  • Test-time scaling: Test-time scaling usually increases depth by extending one trajectory or breadth by sampling multiple trajectories and aggregating them afterward.These approaches include chain-of-thought, structured search, extended thinking, summarization or folding, self-consistency, repeated sampling, and learned aggregators.
  • Test-time scaling: Unlike conventional test-time scaling, AgentFugue exchanges intermediate evidence among rollouts during exploration rather than keeping them mutually opaque.This exchange occurs through the shared reasoning hub.

5 Conclusion

The paper studies scaling out for long-horizon agentic reasoning, asking whether multiple peer trajectories can improve one another on the same task. AgentFugue implements this through a shared reasoning hub that records compact episode notes and supports intent-driven access to teammate trajectories.

  • 5 Conclusion: AgentFugue treats scaling out as a complementary axis to strengthening a single agent trajectory.It evaluates whether multiple peer trajectories can improve one another while solving the same task.
  • 5 Conclusion: The shared reasoning hub writes compact notes from completed episodes and supports intent-driven reading over teammate trajectories.This hub operationalizes communication among peer agents.
  • 5 Conclusion: The study evaluates AgentFugue across homogeneous and heterogeneous teams.The supplied conclusion passage reports results across both team types, but its final claim is truncated.

A Implementation Details

The implementation standardizes tools and interaction budgets across multi-agent systems while specifying AgentFugue’s context, hub-write trigger, and benchmark-dependent evaluation setup.

  • Interaction budget: 150 rounds are allocated per query across systems; Naive- and Swarm-Multi-Agent split this into 100 subagent and 50 meta-agent rounds, whereas each AgentFugue peer receives 150.AgentFugue has no meta-agent, so its N peer agents each share the full per-query cap.
  • Context and hub-write trigger: Each peer agent has a 128k context window, and AgentFugue writes its running state to the hub at 64k tokens before continuing from a compressed prompt.The hub is initialized from Qwen3.5-9B and trained as described in §2.3.
  • Evaluation protocol: Main results follow official benchmark judging: LLM-as-a-judge for BrowseComp and HLE, and structured field-level matching for WideSearch.Exact judges and metrics are reproduced in Appendix B.

B Detailed Benchmark Descriptions … C.3 Multi-Agent Systems

The appendix details three benchmarks, the single-agent and deep-research baselines, and multi-agent comparison systems used to evaluate AgentFugue. The benchmarks span multi-hop browsing, breadth-oriented evidence collection, and closed-book multi-step reasoning, while the baselines vary models, scaffolding, training, and coordination.

  • B Detailed Benchmark Descriptions: BrowseComp requires multi-hop web search, cross-document evidence aggregation, link following, and evidence verification before a short factual answer.Questions are designed so that a single search query cannot reach the answer.
  • B Detailed Benchmark Descriptions: WideSearch emphasizes breadth by requiring agents to collect and consolidate many partially independent pieces of evidence while avoiding omissions.Example tasks include enumerating entity attributes or building structured tables from independent sources.
  • B Detailed Benchmark Descriptions: HLE tests expert-authored questions across mathematics, natural and social sciences, and humanities under closed-book or limited-tool conditions.Its focus is deliberate multi-step reasoning rather than web navigation, probing transfer beyond search-heavy workloads.
  • C Detailed Baseline Descriptions: Baseline descriptions specify each system’s underlying model, agentic scaffolding or training, and instantiation for comparison.This appendix expands the baselines summarized in §3.
  • C.1 LLM-based ReAct Agents: LLM-based ReAct baselines place frontier or open-weight language models in a standard ReAct loop with the same web-search and page-visit tools used by AgentFugue.Closed-source models use official APIs and the same scaffold, while Qwen3.5-35B-A3B provides a like-for-like single-agent reference.
  • C.2 DeepResearch Agents: DeepResearch baselines are single-agent systems combining tool use with long-horizon research scaffolding such as search planning, summary memory, iterative refinement, or web-research post-training.Reported benchmark numbers are taken from the original papers or technical reports rather than rerunning these systems.
  • C.3 Multi-Agent Systems: Multi-agent baselines share AgentFugue’s Qwen3.5-35B-A3B backbone, per-agent context budget, and web tools, differing only in spawning, communication, and timing.This setup isolates coordination mechanisms.
  • C.3 Multi-Agent Systems: Naive-Multi-Agent decomposes a question into K subtasks, runs independent ReAct subagents in parallel, and aggregates their completed reports through a meta-agent.Subagents do not see one another’s progress during execution.

D Answer Aggregation Strategies · E Case Studies: When Shared Memory Helps and When It Misleads

The paper compares lightweight, training-free strategies for combining peer-agent answers and uses case studies to show that shared page memory can either accelerate exploration or induce confirmation bias. These strategies isolate aggregation effects while the case studies illustrate when the shared-memory mechanism helps or misleads.

  • D Answer Aggregation Strategies: Aggregation operates on the same N trajectories, combining each agent’s answer, self-reported confidence, and tool-call count into one team prediction.The restriction is intended to ensure differences reflect output combination rather than additional compute.
  • D Answer Aggregation Strategies: Best-of-N selects the answer from the agent with the highest self-reported confidence and serves as the default aggregator.Self-reported confidence is adopted as a cheap alternative to external reward-model scores.
  • D Answer Aggregation Strategies: Majority Vote selects the most frequent answer, breaking ties by self-reported confidence to measure gains from peer agreement.It provides a self-consistency-style baseline for separating agreement effects from individual confidence calibration.
  • D Answer Aggregation Strategies: Weighted Majority Vote assigns each candidate answer a weight equal to its self-reported confidence and chooses the answer with the largest total weight.WMV interpolates between confidence-only Best-of-N and agreement-only Majority Vote.
  • D Answer Aggregation Strategies: Fewest Tool Calls returns the answer produced with the smallest tool-call count, breaking ties by self-reported confidence.This data-free selector tests the accuracy cost of preferring terse trajectories under the heuristic that shorter correct trajectories are typically more reliable and cheaper.
  • D Answer Aggregation Strategies: Average evaluates agents independently and averages per-agent correctness, measuring the expected accuracy of one sampled agent.It is a reference point: an aggregator that cannot beat Avg does not exploit the team.
  • D Answer Aggregation Strategies: Pass@k counts a question as solved when at least one of k independently sampled agents is correct, with Pass@N matching the Oracle.Values of k below N characterize coverage from parallel-rollout diversity separately from the selection rule.
  • E Case Studies: When Shared Memory Helps and When It Misleads: Two BrowseComp runs with the same 3-agent configuration show shared page memory both accelerating downstream exploration and inducing confirmation bias that overrides hard constraints.Both use the multi_agent_react_with_mem_v2 hub, which compresses evicted live context into shared pages and supplies downstream agents with page summaries.

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

In this puzzle, Agent-0 reaches the gold answer, 1853, after using shared memory as a failure map rather than copying a teammate’s answer. The memory preserves rejected candidates, reasons for rejection, narrowed framing, and open questions, enabling a targeted pivot instead of restarting from scratch.

  • Outcome: Only Agent-0 reaches the gold answer, 1853; Agent-1 ends at 1885 and Agent-2 at 1848.Agent-0 never directly copies a teammate answer, whose consulted pages state that the store remains unidentified.
  • Memory contribution: Agent-0 retrieves a failure map listing store candidates and precise rejection reasons, including “too late,” “no YSB connection,” and “no founder-from-Canton fit.”The candidates include Sincere, Wing On, Sun Sun, Lane Crawford, Whiteaway Laidlaw, and Hall & Holtz.
  • Memory contribution: The memory also records that the store remains unidentified while pointing toward an Eastern-Shanghai foreign-cloth store and Nanjing Road.It preserves related inferences about Yokohama Specie Bank, a Canton founder, and a Chang’an/Xi’an manager.
  • Downstream use: After confirming that the Hall & Holtz/Yokohama Specie Bank lead dead-ends, Agent-0 reuses the narrowed “Eastern Gate, foreign cloth, 1850s” framing to issue a new query.Agent-0 then cross-checks remaining clue mismatches before committing to 1853, rather than restarting from scratch.
  • Take-away: Shared memory transports process-level state—ruled-out directions, rejection reasons, and open sub-problems—rather than answer content.The section identifies this as the regime where collective reasoning can avoid repeatedly revisiting the same dead ends.

E.2 Failure Case: Memory-Induced Confirmation Bias · F Key Prompts

The failure case shows that shared memory preserved disqualifying evidence but amplified a unique clue, causing confirmation bias in an eight-constraint puzzle. The key prompts define page summarization, goal-directed consultation, incremental integration, and memory-page retrieval.

  • E.2 Failure Case: Memory-Induced Confirmation Bias: Despite 10 explicit memory calls, the team answered Texas Prison System Central State Farm instead of the gold answer, Fort Henry, for an eight-constraint puzzle.The puzzle required eight constraints, including enrollment, prison use during two wars, contributor-family affiliation, and city-population conditions.
  • E.2 Failure Case: Memory-Induced Confirmation Bias: Memory contained an explicit rejection ledger and authoritative figures showing Texas A&M enrollment and Sugar Land population fell outside the required ranges.Enrollment was 53,219, 62,137, and 64,326 for 2013–2015; Sugar Land population was 86,972 and 87,367 for 2015–2016.
  • E.2 Failure Case: Memory-Induced Confirmation Bias: The Gieseke/Central State Farm lead remained salient because it was repeatedly recorded as the only clearly confirmed match for the contributor’s father-faculty clue.The same memory also recorded failures on population, growth, and prison-use criteria.
  • E.2 Failure Case: Memory-Induced Confirmation Bias: Agent-2 converted the specific Gieseke clue into a 65%-confidence answer despite acknowledging conflicting or unmet criteria.Its reasoning treated nearby Texas A&M as satisfying the university clue and regarded several mismatches as close, unclear, or conflicting.
  • E.2 Failure Case: Memory-Induced Confirmation Bias: The takeaway is that compressed text memory made local clue uniqueness more salient than hard-constraint conjunctions, while downstream consultation inherited and amplified the anchor.The paper proposes structured candidate states such as ACTIVE, RULED_OUT, and HARD_FAIL as a refinement.
  • F Key Prompts: The key prompts instruct a memory manager to compress prior conversation and tool-use history into concise factual summaries of explored, confirmed, unresolved, and useful information.They explicitly prioritize continuation value and omit chit-chat, stylistic details, and repetition unless task-relevant.
  • F Key Prompts: The consultation prompts make retrieval goal-directed: the assistant extracts directly relevant facts from selected pages, preserves important evidence, and integrates prior results with the current page.The incremental template requires outputting only the updated extracted information and summary.
  • F Key Prompts: The memory tool directs agents to inspect page summaries first, retrieve only directly relevant pages, and use memory to recover or verify details unavailable in the current context.Retrieval accepts up to five integer page numbers.

G Limitations and Broader Impact

AgentFugue’s results are promising but remain limited by narrow model and task coverage, plus failure modes from misleading notes and reduced trajectory diversity. Its peer-sharing approach could benefit knowledge-intensive work, while also amplifying misuse and propagating errors across teams.

  • Limitations: The study evaluates a moderate-sized reasoning hub with limited agent backbones and configurations, leaving transfer across stronger models, families, and hub capacities open.The full space of stronger base models, alternative model families, and larger hub capacities has not yet been evaluated.
  • Limitations: Evaluation focuses on challenging long-horizon reasoning benchmarks, excluding open-ended report writing, sustained software engineering, and richer real-world interactive workflows.The framework may be extensible to these settings, but its behavior there remains unvalidated.
  • Limitations: Low-quality, incomplete, or overconfident notes can propagate misleading conclusions, while repeated reading of salient notes can reduce diversity and cause premature convergence.The paper identifies confidence calibration, diversity-aware reading policies, and adaptive note selection as future directions.
  • Broader Impact: Peer agents sharing intermediate reasoning could support scientific assistance, open-domain research, investigative analysis, and other knowledge-intensive tasks.The proposed perspective treats team-based reasoning as an additional route to agent capability scaling.
  • Broader Impact: Stronger multi-agent systems could enable high-volume surveillance, strategic manipulation, deceptive-content generation, and team-wide propagation of erroneous conclusions.Their abilities to coordinate evidence gathering, synthesize partial discoveries, and scale across agents may amplify misuse.

H Ablation: Hub Context-Window Size … 3. Theory assumptions and proofs

The hub context budget peaks at 32K, while both smaller and larger budgets reduce accuracy; the reported configuration is deliberately conservative rather than an upper bound. The paper also clarifies its LLM usage, claims, limitations, and absence of formal theory.

  • H Ablation: Hub Context-Window Size: At N=2 on BrowseComp, the ablation varies only the per-agent hub context-window budget across five aggregator rules.The aggregators are Pass@N, MV@N, WMV@N, BoN@N, and FewTool@N.
  • H Ablation: Hub Context-Window Size: Accuracy is non-monotone in hub context budget, peaking at 32K and degrading at both extremes across all five aggregators.Very small budgets cut off useful evidence, whereas very large budgets dilute attention and create memory pressure.
  • H Ablation: Hub Context-Window Size: The headline results use a conservative 64K hub-write trigger inside a 128K agent context, not the tuned method’s upper bound.This choice preserves long-trajectory headroom for harder questions.
  • I LLM Usage: LLMs serve as peer agents, the shared reasoning hub, comparison baselines, and LLM-as-a-judge evaluators in the experiments.In paper writing, LLM use is limited to language polishing.
  • NeurIPS Paper Checklist: The abstract and introduction state that the shared hub makes peer-agent count a scaling axis and transfers across homogeneous and heterogeneous teams.The checklist says these claims are substantiated by Table 1 and the scaling studies in §3.4 and §3.5.
  • NeurIPS Paper Checklist: The checklist requires claims to match theoretical and experimental results while reflecting how broadly the findings should generalize.It also emphasizes clearly stating important assumptions and limitations.
  • NeurIPS Paper Checklist: The paper should discuss limitations, strong assumptions, and robustness to violations such as dependence, noise, model misspecification, or locally valid approximations.The checklist encourages a separate Limitations section.
  • 3. Theory assumptions and proofs: The paper is empirical and contains no formal theorems or proofs, so the theoretical-results checklist is not applicable.The checklist explains that NA denotes the absence of theoretical results.

4. Experimental result reproducibility … 14. Crowdsourcing and research with human subjects

The paper documents its architecture, optimization recipe, evaluation settings, and baseline configurations, while noting that compute-cost details and bootstrap confidence intervals remain for the camera-ready version. It reports no human-subjects research or new public assets, uses existing assets under their original terms, and discusses broader impacts and safeguards.

  • 4. Experimental result reproducibility: The paper specifies AgentFugue’s architecture and hub-optimization recipe, with appendices covering tools, budgets, context windows, triggers, benchmarks, judging, and baselines.These disclosures are intended to support reproduction of the main experimental comparisons.
  • 4. Experimental result reproducibility: The experiments report point-estimate accuracy rather than confidence intervals, while scaling and ablation studies characterize variability indirectly.The authors plan to add bootstrap confidence intervals over per-question outcomes in the camera-ready version.
  • 8. Experiments compute resources: Appendix A documents per-agent context windows, interaction budgets, and hub triggers, but detailed training hardware and inference wall-clock costs are deferred to the camera-ready version.The missing details concern training hardware for the hub and inference time per benchmark.
  • 9. Code of ethics: The work uses public benchmarks and pretrained models, releases no personal or sensitive data, involves no human-subjects experimentation, and preserves anonymity.The authors state that they reviewed the NeurIPS Code of Ethics.
  • 9. Code of ethics: The paper discusses positive impacts from more capable assistants and negative risks from coordinated misuse and erroneous hub conclusions, alongside access control, monitoring, and diversity-preserving policies.These safeguards are presented as mitigation directions rather than as released assets.
  • 12. Licenses for existing assets: Third-party benchmarks and pretrained backbones are cited at first use, used under their original licenses and terms, and not redistributed.The cited assets include BrowseComp, WideSearch, HLE, and the listed pretrained backbones.
  • 13. New assets: The submission introduces no new public dataset, model checkpoint, or code package, and the trained hub may be released later but is not included.Consequently, there are no new public assets to document alongside this submission.
  • 14. Crowdsourcing and research with human subjects: The paper involves no crowdsourcing or human-subjects research; all evaluation runs use existing public benchmarks judged by automated protocols.Participant instructions, screenshots, compensation, and IRB details are therefore not applicable.
Loading 2605.24486v1…