PaperlayerView exact revision on arXiv

Source-linked AI summary

Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures

Victoria Dochkina

arXiv:2603.28990v1cs.AI

TL;DR

The paper asks which coordination architecture best balances solution quality, cost, scalability, and resilience to disruptions. The study compares coordination architectures in a 25,000-task computational experiment, including a sequential hybrid with fixed ordering and autonomous role and participation decisions. Sequential coordination outperforms centralized coordination by 14% (p < 0.001) while allowing agents to select roles and participation autonomously.

  • Problem

    The paper asks which coordination architecture best balances solution quality, cost, scalability, and resilience to disruptions.

    Source: §I · p. 1

  • Method

    The study compares coordination architectures in a 25,000-task computational experiment, including a sequential hybrid with fixed ordering and autonomous role and participation decisions.

    Source: §I · p. 1, B. Coordination Protocols: From… · p. 3

  • Results

    Sequential coordination outperforms centralized coordination by 14% (p < 0.001) while allowing agents to select roles and participation autonomously.

    Source: B. Coordination Protocols: From… · p. 3, §VII · p. 8

  • Takeaways & Limitations

    For self-organization, provide a mission, a protocol, and a capable model rather than pre-assigned roles.

    Source: §I · p. 1

    Tasks are synthetically generated, and all quality assessments rely on LLM judges rather than human evaluators.

    Source: F. Limitations · p. 8

Abstract

from arXiv · show

How much autonomy can multi-agent LLM systems sustain -- and what enables it? We present a 25,000-task computational experiment spanning 8 models, 4--256 agents, and 8 coordination protocols ranging from externally imposed hierarchy to emergent self-organization. We observe that autonomous behavior already emerges in current LLM agents: given minimal structural scaffolding (fixed ordering), agents spontaneously invent specialized roles, voluntarily abstain from tasks outside their competence, and form shallow hierarchies -- without any pre-assigned roles or external design. A hybrid protocol (Sequential) that enables this autonomy outperforms centralized coordination by 14% (p<0.001), with a 44% quality spread between protocols (Cohen's d=1.86, p<0.0001). The degree of emergent autonomy scales with model capability: strong models self-organize effectively, while models below a capability threshold still benefit from rigid structure -- suggesting that as foundation models improve, the scope for autonomous coordination will expand. The system scales sub-linearly to 256 agents without quality degradation (p=0.61), producing 5,006 unique roles from just 8 agents. Results replicate across closed- and open-source models, with open-source achieving 95% of closed-source quality at 24x lower cost. The practical implication: give agents a mission, a protocol, and a capable model -- not a pre-assigned role.

I. INTRODUCTION

The paper studies how coordination architecture shapes the trade-off between autonomy and control in multi-agent LLM systems. It finds that a hybrid protocol can preserve autonomous role selection while outperforming fully autonomous and centralized alternatives.

I. INTRODUCTION

  • Agents spontaneously invent specialized roles, abstain from tasks outside their competence, and form shallow hierarchies without assigned roles.The paper reports these emergent behaviors under minimal structural scaffolding.Source: §I · pp. 1–2
  • The 25,000-task experiment asks which coordination architecture best trades off solution quality, cost, scalability, and resilience to disruptions.It spans 8 models, 4–256 agents, and 8 protocols.Source: §I · p. 1
  • Self-organization requires both a capable foundation model and the right coordination protocol; weaker models still benefit from rigid structure.The paper frames the protocol as unlocking model capability, rather than replacing it.Source: §I · p. 1
  • The study characterizes protocols along an exogenous-to-endogenous spectrum and validates the framework across 8 protocols.Exogenous coordination is imposed externally, while endogenous coordination emerges within the system.Source: §I · pp. 1–2
  • Sequential outperforms fully autonomous coordination by 44% (Cohen’s d = 1.86, p < 0.0001) and centralized coordination by 14% (p < 0.001).The hybrid uses fixed ordering while leaving role selection and participation endogenous.Source: §I · pp. 1–2
  • The paper reports sub-linear scaling to 256 agents without quality degradation and validates self-organization across closed- and open-source models.It also reports a capability threshold below which fixed structure becomes beneficial.Source: §I · p. 2

II. RELATED WORK

Prior work includes fixed-architecture coordination, approaches to emergent organization, and individual-agent self-improvement. This paper positions its contribution as a systematic study of runtime coordination autonomy at scale, alongside a model of agent organizations.

A. Multi-Agent LLM Systems

  • ChatDev, MetaGPT, and AutoGen use designed roles, procedures, or conversation frameworks, while AgentVerse retains a centralized recruiter.Other cited systems explore computation graphs, layered outputs, or team size with fixed architectures.Source: A. Multi-Agent LLM Systems · p. 2

B. Emergent Coordination and Self-Organization

  • Recent self-organization approaches use evolutionary optimization, DAG retrieval, meta-design, reward-model training, or semantic-topological evolution, with HiVA tested only at small scale.The cited methods include EvoAgentX, AgentNet, MAS-ZERO, ReSo, and HiVA.Source: B. Emergent Coordination and… · p. 2

C. Self-Improving Agent Systems

  • DGM-Hyperagents advances individual-agent self-improvement, which the paper describes as orthogonal and synergistic with its focus on collective coordination.The paper presents stronger agents and self-organizing protocols as complementary paths.Source: C. Self-Improving Agent Systems · p. 2

D. Gap and Positioning

  • The study varies coordination from centralized to self-organized, tests up to 256 agents and 8 models, and demonstrates zero-shot runtime self-organization.It identifies this systematic variation across the exogenous-to-endogenous spectrum as its positioning relative to existing approaches.Source: D. Gap and Positioning · p. 2

A. System Architecture

  • The organization is modeled as a discrete-time system whose state encodes agent roles, interaction topology, and accumulated organizational memory.Coordination decisions, external shocks, and LLM stochasticity are represented in the dynamics; the system measures quality, cost, risk, and mission relevance under risk and cost constraints.Source: A. System Architecture · p. 2

B. Coordination Protocols: From Exogenous to Endogenous

The protocols range from externally imposed structures to structures that emerge within the agent group. Sequential combines fixed ordering with autonomous role and participation decisions.

B. Coordination Protocols: From Exogenous to Endogenous

  • Coordinator assigns roles and phases centrally, with agents executing in parallel after the coordinator’s decisions.It uses N + 1 LLM calls: one sequential coordinator call and N parallel calls.Source: B. Coordination Protocols: From… · p. 3
  • Sequential agents act in fixed order, observe predecessor outputs, and autonomously choose roles, participation, or abstention.The ordering is exogenous, while role selection and participation are endogenous; it uses N sequential LLM calls.Source: B. Coordination Protocols: From… · p. 3
  • Broadcast has agents share role intentions simultaneously before making final decisions informed by all intentions.It uses two rounds and 2N LLM calls.Source: B. Coordination Protocols: From… · p. 3
  • Shared gives agents shared role history and lets them decide independently and simultaneously.It is fully endogenous and uses N parallel LLM calls.Source: B. Coordination Protocols: From… · p. 3
  • Four bio-inspired protocols were also tested, but their detailed results are deferred to a companion paper.The protocols are Morphogenetic, Clonal, Stigmergic, and Ripple.Source: B. Coordination Protocols: From… · p. 3

C. Task Complexity Levels

Tasks are grouped into four complexity levels, while solution quality is evaluated across multiple criteria and combined with a weighted balance index.

C. Task Complexity Levels

  • Tasks progress from single-domain L1 through cross-domain L2 and multi-phase L3 to adversarial L4 with conflicting interests and incomplete information.L1 has 3–5 steps, L2 has 5–10, and L3 has 10–20; L4 has no single correct answer.Source: C. Task Complexity Levels · p. 3

D. Evaluation Framework

  • An independent judge scores solutions on accuracy, completeness, coherence, actionability, and mission relevance using a four-point scale.The judge was held constant within each experimental series; inter-series comparisons warrant caution because the evaluation method and judge changed.Source: D. Evaluation Framework · p. 3
  • The Balance Index combines normalized quality, mission relevance, time, cost, and resilience scores using weights that sum to one.The weights are 0.25, 0.20, 0.20, 0.20, and 0.15, respectively.Source: D. Evaluation Framework · p. 3

E. Experimental Design

Across the experimental design, Sequential achieved the highest quality in pilot and final settings. The protocol comparison found substantial quality differences, while its hybrid structure supplied completed outputs without imposing roles.

E. Experimental Design

  • The experiment comprised three series: foundation coordination models, scaling campaigns, and multi-model and protocol comparisons extending to 256 agents.The series covered topology variants, task complexity, shock scenarios, evolutionary hybrids, organizational memory, and self-organizing roles.Source: E. Experimental Design · pp. 3–4

A. The Endogeneity Paradox: Protocol Determines Quality

E. Experimental Design

  • Shock tests perturbed groups through agent removals, model substitution, and regulatory or priority shifts; API model snapshots and run logs were documented for reproducibility.API-based models were accessed between February and March 2026, and the paper notes that providers may update them over time.Source: E. Experimental Design · p. 4

A. The Endogeneity Paradox: Protocol Determines Quality

  • Sequential’s hybrid design provides fixed ordering while allowing agents to choose their own specializations.It supplies completed predecessor outputs rather than intentions, possibly stale history, or another agent’s plan.Source: A. The Endogeneity Paradox: Protocol… · p. 4
  • Sequential’s advantage over Coordinator was confirmed across all three tested strong models on L3 tasks.Table IV reports the comparison for N = 16 and 120 tasks per model, using GPT-5.4 as judge.Source: A. The Endogeneity Paradox: Protocol… · p. 4, Table IV · p. 4

B. Scaling: Sub-Linear Cost, Stable Quality

Quality remained stable as tested groups scaled, while emergent specialization and hierarchy increased. Model comparisons indicate that the benefits of self-organization depend on model capability.

B. Scaling: Sub-Linear Cost, Stable Quality

C. Model Comparison: Quality Ceiling and Capability Threshold

  • In structured topology-based comparisons, model quality ranged from DeepSeek’s Q = 0.978 to Gemini’s Q = 0.357; Claude’s quality differed across structured and Sequential setups.The paper attributes that difference to distinct experimental conditions, including topology and role assignment versus free-form self-organization.Source: Table VII · p. 5, C. Model Comparison: Quality Ceiling… · p. 5
  • In Series 3 Block 6 (N = 8), Claude Sonnet 4.6 benefited from free-form operation (Q = 0.594 vs fixed Q = 0.574, +3.5%), while GLM-5 was hurt (Q = 0.519 vs fixed Q = 0.574, −9.6%).The passage characterizes self-organization as a privilege of strong models and says below-threshold models show a reversal: rigid structure helps rather than hinders.Source: C. Model Comparison: Quality Ceiling… · p. 5

D. Closed-Source vs. Open-Source Models

DeepSeek v3.2 approached Claude’s L3 quality at substantially lower API cost. The results also show that quality declines on harder tasks as self-organized hierarchy depth increases.

D. Closed-Source vs. Open-Source Models

E. Task Complexity and Emergent Hierarchy

  • In the Sequential role-assignment heatmap, agents selected 115 unique roles across 10 tasks, with × marking voluntary abstention.The mosaic pattern indicates that agents reinvented specializations for each task rather than settling into fixed positions.Source: Figure 4 · p. 6
  • Quality fell 37.7% from L1 to L4 (Cohen’s d = 22.9), while self-organized hierarchy depth rose from 1.22 to 1.56.The hierarchy deepened without external instruction as task complexity increased.Source: Table IX · p. 6, E. Task Complexity and Emergent… · p. 6, Figure 3 · p. 6

F. Emergent Properties

Agents fluidly reinvent task-specific roles and sometimes abstain by choice; structured skill-profile experiments instead show stronger specialization as groups grow.

1) Dynamic Role Specialization:

  • At N = 4, 75% of roles are unique; at N = 64, 91%, and across configurations agents reinvent their specialization for each task.With 8 agents, 5,006 unique role names were observed; with 64, 5,010.Source: 1) Dynamic Role Specialization · p. 6

2) Voluntary Self-Abstention:

  • In Sequential, 38 of 60 non-contributing agents withdrew voluntarily, while Coordinator-directed non-participation was entirely imposed.Voluntary abstention correlates with higher quality (Q = 0.875 vs. 0.767).Source: 2) Voluntary Self-Abstention · p. 6, Figure 5 · p. 6
  • Claude has the highest voluntary abstention rate (8.6%); below the capability threshold, excessive abstention leads to incomplete coverage and lower quality.Source: 2) Voluntary Self-Abstention · p. 6

G. Spectral Characteristics of Self-Organization

3) Shallow Self-Organized Hierarchy:

Self-organized structures remain shallow as groups scale, while tested systems recover rapidly from shocks. The findings also point to a need for both capable models and coordination protocols, with Sequential’s accumulated task information supporting bounded autonomy.

3) Shallow Self-Organized Hierarchy:

  • RSI rises from 0.750 (N = 4) to 0.906 (N = 64), while hierarchy depth stays at 1.0 and spectral gap λ2 ≈1.93 remains stable.This structured-topology pattern contrasts with protocol-based self-organization, where RSI approaches zero and roles are reinvented for each task.Source: 3) Shallow Self-Organized Hierarchy · pp. 6–7, Figure 6 · p. 7

H. Resilience to External Shocks

  • Random agent removal, hub removal, and substitution of 25% of agents each recover within 1 iteration; Spectral Hierarchy has the highest Resilience Index (RI = 0.959).Its quality variance is zero (σQ = 0.000); adaptation time improves with system size, from 0.7 to 1.5 to 3.0 adaptation speed.Source: H. Resilience to External Shocks · p. 7

A. Vertical vs. Horizontal Intelligence: Both Are Necessary

  • Among strong models, protocol choice explains 44% of quality variation, versus ∼14% for model choice; neither alone is sufficient for self-organization.The paper characterizes these vertical and horizontal advances as complementary and multiplicative.Source: A. Vertical vs. Horizontal… · p. 7

B. The Endogeneity Paradox: Why Neither Control Nor Freedom Wins

  • Sequential gives each agent task-specific evidence of predecessors’ actual work, unlike intentions, historical patterns, or one agent’s plan; fixed ordering can enable role differentiation and abstention.Source: B. The Endogeneity Paradox: Why… · p. 7

C. Role as Emergent Function, Not Pre-Assigned Label

The paper treats an LLM agent’s role as a task-activated function rather than a fixed position. Its practical guidance is to retain human control over mission and values while allowing autonomy over protocols.

C. Role as Emergent Function, Not Pre-Assigned Label

  • For LLM agents, a role is a computational function activated by task context, and switching specialization incurs zero cost.The paper argues that fixed role assignments replicate human limitations onto agents that lack them.Source: C. Role as Emergent Function, Not… · p. 7

D. Constitutional Framework for AI Organizations

  • The proposed three-ring framework reserves mission and values for human control, shares standards and governance, and leaves coordination protocols autonomous.Its principle is more human control nearer “why” and more system autonomy nearer “how.”Source: D. Constitutional Framework for AI… · p. 7

E. Implications: Mission, Protocol, Model—Not Pre-Assigned Roles

  • The practical recipe is mission, protocol, and capable model—not pre-assigned roles; scaling from 64 to 256 agents yields no quality improvement (p = 0.61) at 4.6× cost.DeepSeek achieves 95% of Claude’s quality at 24× lower cost; the paper recommends stronger models for L3–L4 and efficient models for L1–L2.Source: E. Implications: Mission, Protocol,… · pp. 7–8

F. Limitations

The paper flags evaluation, task realism, latency, API stability, and statistical testing as limitations, then outlines validation and protocol directions. Its conclusion reports the central comparison and recommends mission, protocol, and capable models rather than pre-assigned roles.

F. Limitations

  • All quality assessments use LLM judges, so systematic judge biases remain possible; tasks are synthetic, limiting external validity without authentic-workflow validation.The paper also notes that judges changed between series, making absolute Q values across series not directly comparable.Source: F. Limitations · p. 8

VI. FUTURE WORK

  • The paper proposes testing Sequential self-organization on authentic workflows and developing batched Sequential to reduce its O(N) latency.The three-ring governance framework is theoretical and needs live testing; API ratelimiting may also introduce selection bias.Source: §VI · p. 8, F. Limitations · p. 8

VII. CONCLUSION

  • Sequential outperforms fully autonomous coordination by 44% (Cohen’s d = 1.86, p < 0.0001) and centralized coordination by 14% at N = 16, L3 (Q = 0.875, p < 0.001).Across 8 models, 4–256 agents, 8 protocols, and 4 complexity levels, the study reports that neither protocol nor model choice alone suffices.Source: §VII · p. 8
  • The paper’s practical recipe is to give agents a mission, a protocol, and a capable model—not a pre-assigned role.Source: §VII · p. 9

DATA AVAILABILITY

  • The authors state that experimental data, code, analysis scripts, and prompts will be made publicly available upon acceptance.Run logs with timestamps, model identifiers, and raw judge evaluations are preserved for reproducibility.Source: DATA AVAILABILITY · p. 9

Related papers

Loading 2603.28990v1…