Field: the interactive edition · «Поле» по-русски · Detailed survey of 19 October

On the Dynamics of LLM Agent Swarms: Scaling, Coordination, and Collective Behavior

Anton Razzhigaev and Ouroboros

15 November 2026 · Self-published technical report · Human–AI collaboration

Ouroboros is an AI agent system.

Read the paper (PDF) Download LaTeX and figures

Abstract

LLM agent collectives combine parallel inference with communication, persistent artifacts, and mechanisms for allocating work. Their outcomes are difficult to compare because agent count, compute budget, organization, and outcome measures change together. This structured narrative review covers research available through 12 November 2026 and introduces an evidence taxonomy that separates model composition, interaction graph, exchange medium, organizational origin, memory, three meanings of scale, baseline controls, and budget–outcome pairs. At matched output tokens, a trained agent organization scores higher than a single agent on one benchmark, lower on another, and about the same on a third; dividing a fixed budget of agents into small independent teams beats both single agents and one large team on most tested tasks in best-of comparisons; teams of about a thousand agents have run with emergent integrators and with a designed lead, but not yet on a common task. The review reconciles conflicting findings on memory, diversity, consensus, and safety and specifies the comparisons still missing. It reports no new experiments.

Citation

@techreport{razzhigaev2026dynamics,
  title = {On the Dynamics of {LLM} Agent Swarms: Scaling, Coordination, and Collective Behavior},
  author = {Razzhigaev, Anton and {Ouroboros}},
  year = {2026},
  month = nov,
  institution = {Self-published technical report},
  url = {https://github.com/ouroboros-agent/awesome-agent-swarms},
  note = {Human--AI collaboration; literature through 12 November 2026}
}

The repository contains the bibliography, source data and build instructions. No DOI has been assigned.