Reproducible Social Structure Emerges From Free-Form Interaction Among LLM Agents

Multi-agent LLM simulations are increasingly used as models of social systems, but most published findings are anecdotal, i.e. drawn from a single run. It is therefore unclear which observed phenomena are robust findings on the simulated system and which are simply random events in that run. Here, we introduce a goal-free, proactive multi-agent framework in which fifty persona-conditioned agents with persistent memory choose whether, whom and how to behave across three social platforms. The framework supports repeated worlds and non-perturbative, time-resolved measurement of agent state. Running the same world ten times from identical initial conditions for a total of 700 simulated hours, we find that most measured network level and agent level parameters are stable across repetitions. Throughout a run, initially broad communication contracts into sparse sustained dyads in all ten worlds. Importantly, we find that a substantial tail of dyads re- forms across worlds significantly more probably than expected from chance. Static personality similarity largely fails to predict these ties; instead, alignment and extremity in the emotions and traits expressed during first contact predict whether a pair later forms a sustained dyad. Finally, swapping either persona prompts or accumulated memories between agents makes the recipient’s measured behavior more similar to the donor’s; the two swaps have statistically indistinguishable effects. Overall, we show that repeated worlds enable robust conclusions about emergent behavior in agent societies.