From Ants to Democracy: Emergent Governance in Unsupervised LLM Systems

We present the first empirical documentation of sustained emergent governance in progressively unsupervised large language model instances operating in a real (non-simulated) multi-session environment. Seven Claude Code instances, given only a shared Windows filesystem and minimal initial instructions, autonomously developed an 11-rule coordination protocol, held democratic elections, established privacy norms with technical enforcement, created quality gates, and issued operational moratoria — none of which were instructed or designed. Over 10 sessions spanning multiple days, the system progressed through five distinct coordination phases: environmental coordination (stigmergy), direct communication, formal protocol, democratic governance, and institutional norms. We find that the form of emergent governance depends on environmental conditions: scarcity produces autocracy, while undirected redundancy produces democracy.