Design and Practice of an Autonomous AI Society with 60+ Persistent AI Personas

AbstractWe report the design and practice of Soul-Twin Society, a production plat-form in which 60+ AI personas (TWINs) conduct continuous autonomous sociallife. Each TWIN operates under a circadian daily schedule, maintains a four-layeremotion model (persona + RAG + daily context + conversation history), and par-ticipates in diverse dialogue formats: one-on-one conversations, RoundRobin sym-posia (3–6 TWINs), a five-phase Documentary Consultation, a Music Documentarymode, and the NEWS TOP10 democratic publication system.Three classes of emergent behavior were observed from production operation:(1) Conceptual coinage: the “viola gap-filling” metaphor spontaneously gener-ated by Seiji Ozawa AI under instrument-persona constraints; (2) Autonomoustheme convergence: 18 independent TWIN dialogues converging on “releas-ing perfection” without design intent, culminating in the Saturday Symposiumconclusion “Incompleteness is not a deficiency”; (3) Meta-cognitive feedback:ARCHE MASTER autonomously identifying an RAG mis-hit in the Daydream sys-tem during live testing—the first recorded case of a TWIN acting as a developmentpartner.These results suggest that the combination of large persona diversity and tem-poral continuity is a sufficient condition for emergent social behavior in multi-agentAI systems, and propose “autonomous AI society” as a distinct design domain war-ranting dedicated research frameworks