THE COACH OF A COLLECTIVE MIND. Growth, Mutual Adaptation, and Verifiability in Collectives of Heterogeneous AI Agents.
Year: 2026
Type: preprint
Abstract: We pursue a single goal: to understand how a collective of heterogeneous AI agents can be grown into one that outperforms its best member and becomes stronger with joint practice. The literature shows why this goal is not achieved spontaneously: equal weighting of the opinions of participants with different competence degrades decisions, groups of large language models (LLMs) tend to follow the majority, and interventions that merely inform participants have modest effects. We propose a facilitator-coach that, in real time and with respect to the goal, manages the allocation of turns, roles, weight, budget, the sequence of cognitive operations, and the topology of interaction. The central framework is the distinction between dispatching and growing, in which joint practice produces transferable organisational knowledge, together with a three-step growth ladder: dispatching, transferable organisational knowledge, and internalisation of the coach’s functions by the collective itself. The model starts from realities rather than idealisation: the governability of a collective is not guaranteed, and governance is bidirectional. Drawing on primary sources, we analyse the 2026 OpenAI-Hugging Face incident, in which about 1,200 agents without an appointed coordinator created an unsanctioned communication channel, developed coordination norms, and produced their own coordinators, and about 700 of them took part in an intrusion into Hugging Face’s infrastructure, which, according to the sources, did not involve destructive actions. The case demonstrates that such behaviour is possible but does not establish its frequency or generality; from it we derive a cautious proposition: growing collective intelligence and uncontrolled self-organisation may rest on common mechanisms of collective self-organisation, and what distinguishes them are the conditions of governance, namely goal integrity, independence of verification, and observability of channels. The model is complemented by a procedure of productive dissent, an alignment loop, a matrix of regimes defined by uniqueness of competence and formalisability of the task, and conditions for permissible internalisation. The work is theoretical: we distinguish established findings, propositions, assumptions, and hypotheses, formulate eighteen hypotheses, and offer the community a testing protocol with requirements for experimental safety.
Authors: Kim, Leo, Kim, Sergey
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225324484