Organizations: School of Mathematical Sciences, University of Chinese Academy of Sciences · Department of Statistics and Data Science, National University of Singapore · Department of Statistics and Data Science, Tsinghua University
Multi-agent systems (MAS) combine multiple agents to solve complex tasks, and recent MAS scale to more than a thousand agents. At this scale, decentralized MAS remove the central orchestrator that assigns sub-tasks, and the agents, which we call workers, self-organize through their own choices of sub-tasks and collaborators on a shared board. As workers keep adding participation and collaboration records, however, the information explodes: workers spend more and more effort on reading the records of other workers, and self-organization loses efficiency. In the statistical view, we represent the execution history of an MAS by two continually updated networks, a collaboration network among workers and a participation network between workers and task types. We then propose AgentCue, a method built on a new latent space model that places workers and task types in a shared latent space. Each time the networks are updated, the estimated parameters of the latent space model form a compact and interpretable statistical summary, and workers use the statistical summary to choose sub-tasks and collaborators without repeated searches of the shared board. We establish the identifiability of the latent space model and non-asymptotic error bounds for the estimates under network increments, and we show that a fixed number of gradient updates after each network update suffices for the same error rate. Experiments on LongBench v2 Multi-Doc QA show that AgentCue achieves competitive effectiveness, efficiency, and interpretability.
Figures & tables
Figure 1: Self-organization in a decentralized MAS. Workers read from and publish to the shared board (dashed lines) and collaborate with one another (solid lines), and groups of collaborators form and change over time through local choices. With few workers (upper panel), the shared board holds few records. With many workers (middle panel), records accumulate quickly, and each local choice (lower panel), which sub-task to claim and whom to collaborate with, requires a search through many records.
Figure 2: Overview of AgentCue. Upper and middle bands: once per checkpoint, the MAS publishes the statistical summary, and worker i reads the statistical summary once. Lower band: the local comparison, enlarged. Bars are log fitted means split as in ( 4 ); the struck-through term cancels, and hatched segments are negative affinity terms. The latent-space map shows the projections onto the line of sight of worker i . Small squares scale with the fitted participation mean of each candidate in the task type of the inputs that the sub-task needs; the dashed box and the rings mark the shortlist; the magnifier marks a work-record search. Superscripts (t) are omitted.
Figure 3: Simulation of ( 8 ) in the setting of Section 3.4 ; each line is one run. Dotted lines: fixed point without the statistical summary ( ω=0 ). Dashed lines: interior fixed point, stable for ω=0.5 and unstable for ω=2 ; for ω=1 , no interior fixed point exists.
Measure
AgentCue
History only
Correct answers
76.0±1.0
72.0±1.0
Accuracy (%)
60.8±0.8
57.6±0.8
Table 1: Final results over three paired replicates (mean ± sample standard deviation), with 256 effective sessions per trajectory and 125 questions.
Figure 4: Effectiveness, efficiency, and interpretability of AgentCue. (a) Final accuracy by domain; parentheses give question counts. (b) Accuracy over the common effective-session budget, using all 125 questions. Points and lines show means across three paired replicates; error bars and shading show one sample standard deviation. (c) One recorded choice in R1 links the fitted participation means of W15, the choice of inference , the use of version 1 of the published result of W26, and the correct designated answer, where W i denotes worker i .
Figure 5: Worker sessions and statistical summaries for the AgentCue trajectory in R1. Upper panels show eight statistical summaries jointly aligned to the statistical summary at the final checkpoint by orthogonal Procrustes, with consecutive checkpoints paired by column. Linear axes show worker latent positions; insets enlarge the four task-type latent positions. The color scale links panels and arrows to the checkpoints t1,…,t8 . Lower bars show all 256 sessions of 32 workers, mapped from timestamps to the number of completed effective sessions. Equal-height stripes indicate task types of work published in a session; gray denotes no publication.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Task type
Required independently reusable result
Extract
Source-located facts, values, conditions, or a documented search result.
Compare
Named objects, shared dimensions, and evidence-supported similarities, differences, or conflicts.
Inference
Inputs or premises, reasoning, conclusion, and applicable conditions. Answer synthesis also gives an option and explains its use of published results of other workers.
Review
The exact version of another worker’s published result, checks, findings, evidence, and a PASS or FAIL verdict.
Appendix
Table \thechapter .B.1: Task types defined by independently reusable results.
Replicate
AgentCue
History only
Difference
PASS
Correct + PASS
Correct
Accuracy (%)
Correct
Accuracy (%)
R1
76
60.8
71
56.8
4.0
36/36
22/20
R2
77
61.6
72
57.6
4.0
33/41
18/23
R3
75
60.0
73
58.4
1.6
92/47
55/24
Appendix
Table \thechapter .B.2: Final results by replicate. Difference is in percentage points; PASS columns count designated answers with a PASS verdict (AgentCue/History only).
Measure
AgentCue
History only
Published works
359.7±26.8
348.0±2.6
Extract
160.7±3.2
162.3±11.2
Compare
140.0±18.4
153.3±6.4
Inference
246.0±56.0
250.7±51.4
Review
71.3±31.0
57.0±5.3
Collaboration count
284.7±46.9
248.7±3.2
Appendix
Table \thechapter .B.3: Participation and collaboration (mean ± sample standard deviation). Task-type counts are column sums of B(T) ; collaboration is ∑i<ℓAiℓ(T) .
Replicate
Method
Works
E
C
I
R
Workers E/C/I/R
Collaboration
R1
AgentCue
350
163
119
290
56
32/32/32/28
274
R1
History only
349
175
156
288
51
32/32/32/25
250
R2
AgentCue
339
162
153
183
51
32/32/32/27
244
R2
History only
345
154
146
192
59
32/32/32/29
251
R3
AgentCue
390
157
148
265
107
32/31/32/31
336
R3
History only
350
158
158
272
61
32/32/32/30
245
Appendix
Table \thechapter .B.4: Participation by replicate. E/C/I/R denote extract/compare/inference/review; the worker column counts contributors to each task type.
Method
Estimates
Iterations
Gradient norm
Z
W
β
AgentCue
24
111–469
2.33×10−6 – 9.95×10−6
1.33×10−15
1.11×10−16
6.22×10−15
Appendix
Table \thechapter .B.5 : Estimation diagnostics over eight checkpoints. Centering columns give maximum absolute sums of latent-position columns or node effects.
Measure
AgentCue
History only
Input tokens (millions)
415.97±2.04
390.06±2.93
Cached input (millions)
392.85±1.91
367.76±2.92
Output tokens (millions)
2.05±0.04
1.94±0.04
Reasoning output (millions)
0.60±0.01
0.56±0.02
Post-warm-up worker span (minutes)
29.81±0.70
24.33±3.35
Additional estimation time (seconds)
0.314±0.021
–
Appendix
Table \thechapter .B.6: Computational resources (mean ± sample standard deviation). Tokens include warm-up; worker spans exclude it. Additional estimation time covers checkpoints 64–256; History only involves no estimation.
A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.
Multi-agent systems (MAS) have recently emerged as an effective approach for coordinating large language model (LLM)-based agents to solve complex tasks through structured interactions. In practice, MASs often handle a stream of heterogeneous and complex tasks, requiring agents to decompose each task and then self-organize and self-evolve to adapt to incoming tasks while sharing execution resources. However, most early approaches to MASs rely on centralized controllers or fixed coordination patterns, which can limit scalability or adaptability. In contrast, existing decentralized and dynamic MASs often require training dedicated routers or invoking LLMs for agent selection, resulting in substantial computational costs and coordination overhead. To address these challenges and enable efficient task-level self-organization and self-evolution for task- and workload-level collaboration, we propose Decentralized Expertise-Aware Load Serving (DEALS), a decentralized and low-complexity framework that enables agents to self-organize and dynamically route concurrent tasks for processing. Specifically, each agent maintains local queues of incoming tasks, and its router decides whether to process a task locally or forward it to a neighbor based on differences in backlog and success rate. Meanwhile, executors process independent tasks concurrently within and across agents, and partially solved tasks can be resumed by other agents. Experiments show that DEALS not only improves performance along multiple dimensions (e.g., answer accuracy and task throughput) in both homogeneous and heterogeneous agent pools, but also balances agent expertise and workload in a self-organized manner, enabling effective decentralized coordination.
Jingjuan Huang, Wenbin Wang, Yanchuan Yin +2
Ohio State University · University of Colorado at Boulder
Self-evolving multi-agent systems (MAS) have emerged as a promising route to LLM agents that continually improve from experience, with persistent memory at their foundation. However, existing designs almost exclusively adopt a centralized repository shared across agents, incurring communication and coordination overhead, raising privacy concerns, and collapsing agent diversity. We propose DecentMem, a decentralized memory framework in which each agent maintains its own dual-pool memory -- an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidates for unseen contexts. The two pools are reweighted online based on stage-wise feedback from an LLM-as-a-judge. Theoretically, we prove that this design guarantees global reachability of the solution space and achieves O(logT) cumulative regret, matching the stochastic bandit lower bound up to constants. In practice, across three MAS frameworks (AutoGen, DyLAN, AgentNet), three Qwen3 backbones (4B/8B/14B), two Gemma4 backbones (E2B/E4B) and five benchmarks spanning math, code, QA, and embodied tasks, DecentMem improves average accuracy by up to 23.8% over the strongest centralized memory baseline and by up to 52.5% over the no-memory baseline, while reducing token usage by up to 49%.