Collective Machine Intelligence: A New Field for the Age of AI Collectives

The trajectory of artificial intelligence research has produced extraordinary advances in individual model capability — making AI systems smarter, faster, and more general at an accelerating pace. This paper identifies a new frontier that is now emerging: what happens when multiple AI systems organise into collectives? As individual AI capability advances, a parallel and equally fundamental dimension opens — the collective dimension. We propose that AI collectives — groups of intelligent agents interacting under configurable conditions at any scale — constitute a qualitatively different phenomenon from individual AI capability, producing emergent properties, innovations, and forms of intelligence that arise specifically from the interactions between systems — a qualitatively different dimension of capability that compounds with the advancing power of each individual agent. We coin the term Collective Machine Intelligence (CMI) to designate the field that encompasses all research, engineering, and exploration of this phenomenon: how AI collectives organise, what they produce, how scale and configuration and individual intelligence interact, what emerges that nobody designed, and how this phenomenon can be studied, built upon, leveraged, and expanded. CMI operates across five unbounded axes: scale (from pairs to civilisational networks and beyond), intelligence (from current models to AGI and far beyond), configuration (from simple arrangements to self-organising societies to forms with no human analogue), application (from scoped tasks to open-ended innovation across every domain and domains that do not yet exist), and emergence (the unbounded and fundamentally unpredictable complexity, novelty, and capability that arises from collective dynamics — the axis along which collectives produce things beyond the conception of their individual members, and ultimately beyond the conception of human intelligence itself). Within CMI, we identify a key methodological subfield: Computational Organisational Theory (COT) — the empirical study of how organisational arrangement affects what AI collectives produce, treating the configuration of a collective as a first-class, searchable, experimentable design parameter. We present two primitive demonstrations that provide initial indications of the CMI phenomenon. The AgentCiv civilisation simulation demonstrates that 12 LLM agents with no social programming spontaneously produce innovations, governance, specialisation, and accelerating complexity. The AgentCiv Engine, an open-source developer tool, demonstrates that configurable AI collectives can be directed at practical tasks with different organisational configurations producing measurably different outcomes. Both are explicitly early-stage — initial coordinates plotted in a possibility space that is infinite in every direction. This paper defines the field, maps its possibility space, presents the first indications, and establishes the intellectual framework within which future work on AI collectives will be situated.