Self-organization of LLM agents without roles +14% quality

LLM Agent Self-Organization: Dropping Roles Boosts Efficiency by 14%

An LLM agent system without rigid roles and hierarchies outperforms traditional coordinator-based approaches by 14% on the quality metric Q (p < 0.001). An experiment on 25,000 tasks with 8 models and up to 256 agents showed that agents independently determine specializations, decline unnecessary tasks, and scale without performance loss. The sequential protocol, where agents work in sequence with access to their predecessors' results, delivered the highest performance.

Experiment Scale and Methodology

The experiment covered over 25,000 tasks across four complexity levels: from API verification (L1) to adversarial scenarios like CEO vs Legal vs CFO (L4). Eight models were tested — Claude, GPT-5.4, GPT-4o, GPT-4.1-mini, DeepSeek v3.2, GLM-5, Gemini-3-flash, GigaChat 2 Max. The number of agents varied from 4 to 256, with total token consumption exceeding 1 billion.

Solution quality was evaluated by a judge model across five criteria: accuracy, completeness, coherence, applicability, and mission alignment. The Q metric ranged from 0.25 to 1.0.

Eight coordination protocols were compared, including four bio-inspired ones. The focus was on horizontal coordination: how a group of agents forms a unified system without vertical reinforcement of individual instances.

Coordination Protocols and Their Effectiveness

Four key protocols were tested with identical models, tasks, and agent counts:

• Coordinator: A central Agent-0 assigns roles; others execute in parallel.
• Sequential: Agents work in turn, each analyzing specific results from previous agents and deciding on role and participation.
• Broadcast: Agents announce intentions, then adjust based on others' announcements.
• Shared: Complete independence with shared memory.

Results on GPT-4.1-mini (N=8):

| Protocol | Quality Q |

|--------------|-----------|

| Sequential | 0.724 |

| Coordinator | 0.640 |

| Broadcast | 0.510 |

| Shared | 0.503 |

The difference between best and worst was 44% (Cohen’s d = 1.86). At N=16 on Claude for complex tasks, Sequential achieved Q=0.875 vs. 0.767 for Coordinator (+14%, p < 0.001).

Sequential wins due to access to actual completed results, not intentions or general histories.

Role and Specialization Dynamics

Eight agents on hundreds of tasks generated 5006 unique roles, 54% of which were used only once. With 64 agents — 5010 roles (a 0.1% difference). Agents adapt their functions to the context of each task, ignoring fixed profiles.

In Sequential, roles change dynamically, forming a distributed network. In Coordinator, all connections pass through Agent-0 — a single point of failure.

Self-Removal and Resource Optimization

Agents in Sequential voluntarily decline tasks: "All key points are covered, I won't add value." Out of 60 inactive agents, 38 were self-removals. This boosts Q to 0.875 vs. 0.767 with forced participation.

Claude shows 8.6% conscious refusals — an optimal level. The system itself determines the needed number of participants, minimizing tokens.

When scaling:

| N agents | Q | Tokens |

|----------|--------|----------|

| 8 | 0.954 | 3164 |

| 64 | 0.949 | 3537 |

| 256 | 0.967 | — |

Cost increased by 11.8% with an 8x increase in N, with ~45% of agents self-removing.

Model Comparison and Cost

On complex tasks:

• Claude Sonnet 4.6: Q=0.875, high cost.
• DeepSeek v3.2: Q=0.829 (~95% of Claude), cost ~1/24.
• GLM-5: Q=0.800, ~1/20.

DeepSeek outperforms Claude on adversarial tasks (+6.0%), with mission alignment 4.00/4.00.

Trend: Strong models (Claude, DeepSeek) benefit from autonomy (+3.5–10.6%). Key properties:

• Reasoning — chain-of-thought capabilities.
• Self-reflection — competence assessment.
• Instruction following — protocol adherence.
• Structured output — stable formatting.

Adaptation to Task Complexity

Q declines from L1 to L4 (-37.7%), but hierarchy spontaneously deepens (1.22 → 1.56). Agents themselves build structure without instructions.

Key Takeaways

• The sequential protocol increases Q by 14% due to access to actual predecessor results.
• Agents generate thousands of unique roles, adapting to context without fixed profiles.
• Voluntary self-removal optimizes resources: 45% of agents don't participate at N=256.
• DeepSeek v3.2 achieves 95% of Claude's quality at 1/24 the cost.
• Autonomy is effective with strong models featuring advanced reasoning and self-reflection.

— Editorial Team