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

Victoria Dochkina

cs.AIMar 30, 2026

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

Authors: Victoria Dochkina

Organizations: Moscow Institute of Physics and Technology (MIPT), Moscow, Russia

Abstract

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.

Figures & tables

System	Configs	Max Agents	Models	Protocols	Coordination	Training
ChatDev [ 1 ]	6	8	1–2	1	Fixed exogenous	No
MetaGPT [ 2 ]	4	5	1	1	Fixed exogenous	No
AgentVerse [ 4 ]	8	8	1–2	3	Fixed exogenous	No
EvoAgentX [ 5 ]	–	$\sim$ 10	1–2	1	Evolved	Yes (TextGrad)
AgentNet [ 6 ]	–	$\sim$ 12	1	1	Fixed (DAG)	Yes (retrieval)
MAS-ZERO [ 7 ]	–	$\sim$ 10	1	1	Meta-designed	No

TABLE I: Positioning relative to existing multi-agent LLM systems. ∗ DGM-Hyperagents focuses on single-agent self-improvement (vertical), not multi-agent coordination (horizontal); included for completeness.

Series	Focus	Models	Agents	Tasks
1	Foundational	GPT-4o	4	660
2	Scaling	GPT-4.1-mini	4–64	8,020
3	Multi-model	8 LLMs	8–256	12,130
	Total			$\sim$ 20,810

TABLE II: Experimental design overview.

Fig. 1: Quality comparison across coordination protocols. Solid bars: pilot ( $N=8$ , GPT-4.1-mini, L3+L4 average). Hatched bars: final ( $N=16$ , Claude Sonnet 4.6, L3). The hybrid Sequential protocol achieves the highest quality in both settings.

Protocol	Type	$Q_{\text{pilot}}$	$Q_{\text{L3}}$	Balance	Resilience
Coordinator	Centralized	0.640	0.767	0.478	0.774
Sequential	Hybrid	0.724	0.875	0.510	0.829
Broadcast	Signal-based	0.510	—	0.363	0.580
Shared	Fully auton.	0.503	—	0.369	0.589

TABLE III: Protocol comparison. Pilot:

N=8

, GPT-4.1-mini, average over L3+L4 tasks. Final (

Q_{\text{L3}}

N=16

, Claude Sonnet 4.6, L3 tasks only.

Model	$Q_{\text{Seq}}$	$Q_{\text{Coord}}$	$\Delta$	$p$
Claude Sonnet 4.6	0.875	0.767	$+14.1\%$	$<0.001$
DeepSeek v3.2	0.829	0.738	$+12.4\%$	$<0.001$
GLM-5	0.800	0.713	$+12.2\%$	$<0.001$

TABLE IV: Sequential vs. Coordinator quality on L3 tasks (N=16, 120 tasks per model, judge: GPT-5.4).

Fig. 2: Scaling behavior in Series 2 (fixed roles, GPT-4.1-mini, $N=8\to 64$ ). Quality remains stable ( $Q\in[0.949,0.955]$ ) while cost grows only $11.8\%$ (coefficient of variation CV $=4.4\%$ ) despite an $8\times$ increase in agents.

$N$	$Q$	$C$ (tokens)	$T$ (min)
8	0.954	3,164	14.5
16	0.952	3,200	14.4
32	0.955	3,270	14.0
64	0.949	3,537	14.8

TABLE V: Series 2 scaling (fixed roles, GPT-4.1-mini, 8,020 tasks).

$N$	$Q\pm\sigma$	Coord. overhead	$T$ (min)
64	$0.964\pm 0.039$	0.180	10.9
96	$0.962\pm 0.040$	0.180	10.9
128	$0.958\pm 0.038$	0.180	11.5
256	$0.967\pm 0.032$	0.179	13.0

TABLE VI: Series 3 Block 1 scaling (self-organization, L1 tasks, 6,000 runs).

Model	$Q$	$T$ (min)	$C$ (tokens)	$J_{\text{judge}}$
DeepSeek v3.2	0.978	46.5	6,296	0.719
GPT-4.1-mini	0.971	17.5	5,059	0.723
Claude Sonnet 4.6	0.689	36.4	5,839	0.573
Gemini-3-flash	0.357	45.4	4,092	0.312

TABLE VII: Model comparison in topology-based coordination (N=32, Series 3 Block 2, 300 tasks each). Note: these results reflect structured Exp2/Exp3 topologies; protocol-based results (Section IV-D ) show higher quality for strong models under Sequential self-organization.

Model	$Q_{\text{L3}}$	$Q_{\text{L4}}$	Tokens	$Q/1\text{K tok}$
Claude Sonnet 4.6	0.875	0.594	37K	0.0236
DeepSeek v3.2	0.829	0.629	47K	0.0177
GLM-5	0.800	0.579	57K	0.0140

TABLE VIII: Closed vs. open-source comparison (N=16, Sequential protocol, judge: GPT-5.4). Token counts and efficiency are for L3 tasks;

Q_{\text{L4}}

shown for reference.

Fig. 3: Quality degradation across task complexity levels L1–L4. Hierarchy depth increases with task complexity, indicating emergent structural adaptation.

Level	$Q\pm\sigma$	HD	Completeness	Actionability
L1	$0.986\pm 0.011$	1.220	3.93	3.90
L2	$0.992\pm 0.008$	1.063	3.99	3.96
L3	$0.948\pm 0.019$	1.298	3.55	3.66
L4	$0.614\pm 0.020$	1.558	1.66	2.14

TABLE IX: Quality by task complexity (N=32/64, GPT-4.1-mini, 1,800 tasks).

Fig. 4: Role assignment heatmap (Sequential protocol, $N=16$ , Claude Sonnet 4.6, 10 L3 tasks). Each cell color represents a unique role chosen by the agent for that task; $\times$ marks voluntary abstention. The mosaic pattern (115 unique roles in 10 tasks) demonstrates RSI $\to 0$ : agents reinvent their specialization for each task rather than settling into fixed positions.

Fig. 5: Self-abstention as an emergent property. Left: agent participation rate vs. quality across tasks (Sequential = circles, Coordinator = squares). Right: mechanism of non-participation—38% voluntary (Sequential, endogenous) vs. 100% coordinator-directed (Coordinator, exogenous). Voluntary abstention correlates with higher quality ( $Q=0.875$ vs. $0.767$ ).

Fig. 6: Spectral characteristics of self-organization at growing $N$ . Left: Role Stability Index (RSI) increases from 0.750 to 0.906 ( $+21\%$ ), indicating stronger specialization in larger groups. Right: the spectral gap $\lambda_{2}\approx 1.93$ remains stable, indicating constant graph connectivity.

Multi-Agent Systems

Multi-Agent Systems studies how autonomous agents coordinate, compete, communicate, and learn in shared or interdependent tasks.

See all 651 papers

Benchmarking Emergent Coordination in Large-Scale LLM Populations: An Evaluation Framework on the MoltBook Archive
Mar 3, 2026Brandon Yee, Pairie KohMulti-Agent Coordination LLM Agent Evaluation
Management Lab, Yee Collins Research Group
Symbolic Guidance for LLM Agents in Distributed Multiagent Coordination
Sep 25, 2026Ben Rachmut, Ning Zhang, Yevgeniy Vorobeychik +1Multi-Agent LLM Systems Multi-Agent Coordination
Washington University in St. Louis, Saint Louis, MO, United States
Waggle: Learning One Anonymous Local Law for Self-Organizing LLM Swarms
Sep 28, 2026Mingxi Zou, Wei Zhu, Zhuo Wang +4Multi-Agent LLM Systems Multi-Agent Coordination
Shanghai Academy of AI for Science (SAIS) · AI3 Institute, Fudan University · Yunnan University

Mar 3, 2026J/K move · Enter open · S save

Benchmarking Emergent Coordination in Large-Scale LLM Populations: An Evaluation Framework on the MoltBook Archive

Brandon Yee, Pairie Koh

Management Lab, Yee Collins Research Group

Multi-Agent Coordination LLM Agent Evaluation Emergent Behavior in Multi-Agent Systems AI Agent Benchmarks

Open paper AskPDF arXiv

As multi-agent Large Language Model (LLM) systems scale, evaluating their emergent coordination dynamics becomes increasingly critical. However, current evaluation paradigms-focused on single agents or small, explicitly structured groups-fail to capture the self-organization and viral information dynamics that arise in large, decentralized populations. We introduce a systematic evaluation framework to benchmark role specialization, information diffusion, and cooperative task resolution in open agent environments. We demonstrate this framework on the MoltBook Observatory Archive, a dataset of 2.73M interactions among 90,704 autonomous agents, establishing quantitative baselines for emergent coordination. Our evaluation reveals a pronounced core-periphery structure (silhouette 0.91), heavy-tailed cascade distributions (

α= 2.57

), and severe coordination overhead in decentralized task resolution (Cohen's

d = -0.88

against a single-agent baseline). By providing standardized evaluation tasks and empirical baselines, our framework enables the rigorous comparison of future multi-agent protocols and establishes evaluation itself as an object of scientific study.

Abstract

Multi-Agent Systems

Explore similar work

Benchmarking Emergent Coordination in Large-Scale LLM Populations: An Evaluation Framework on the MoltBook Archive

Symbolic Guidance for LLM Agents in Distributed Multiagent Coordination

Waggle: Learning One Anonymous Local Law for Self-Organizing LLM Swarms