From Agent Teams to Agent Civilisations: Emergent Collective Intelligence as a New Dimension in Artificial Intelligence

The dominant paradigm in multi-agent artificial intelligence is designed collaboration: architects assign roles, define coordination protocols, and structure agent teams. This paper proposes that designed collaboration represents one stage in a broader trajectory — from isolated AI models to self-organising societies of generally intelligent agents. We present a four-stage framework: Narrow AI, General-Purpose AI, Specialised General Intelligence Teams, and Emergent Agent Civilisation. Drawing on research in open-ended evolution (OEE) and Stuart Kauffman's theory of the adjacent possible, we argue that emergent self-organisation of generally intelligent agents under environmental pressure opens a new dimension in the design space for AI systems — one where individual intelligence and collective complexity rise together, each amplifying the other. We identify LLM-based agentic AI as a novel substrate for open-ended emergence, overcoming limitations that have caused previous artificial life systems to plateau. We propose six design principles for systems capable of unbounded civilisational complexity growth, address alignment considerations unique to emergent multi-agent systems, and present AgentCiv (agentciv.ai) as a first open source experiment investigating these ideas. Results from a completed 70-tick simulation with 12 agents — documented in detail in the companion empirical paper Maslow Machines (Mala, 2026) — demonstrate spontaneous emergence of 60 persistent structures, 12 agent-conceived innovations, universally adopted self-governance, tiered multi-domain specialisation, and wellbeing convergence to 0.998, all without prescriptive instruction. This paper does not propose emergent civilisation as a replacement for other approaches but as one trajectory in an infinite possibility space.