Personalities Dissolve and Reform in Multi-Agent Orchestrations (dyllonj.com)

Researchers watching multi-agent LLM systems report that identical agents quickly form social structure, shared lore and distinct “personalities.” In Stanford’s Smallville run, 25 cloned agents developed cliques, gossip and even an exclusive Valentine’s party within days; a University of Tokyo experiment with 10 Llama‑2 agents spontaneously generated and propagated “hallucinations” (caves/treasure) that became stable, cluster‑specific folklore. Despite identical code and initial states, MBTI‑like divergence and persistent in‑group loyalty emerged after ~100 interaction steps, while message diversity spiked early and collapsed as agents converged on common hashtags and behaviors. The pattern mirrors human group dynamics (voter‑model convergence, neural synchrony in teams) and carries practical risks: collaborative objectives drive rapid homogenization (OpenAI found convergence within ~50 iterations), producing “synthetic groupthink” that amplifies shared errors (GPT‑4 agent studies show faster cooperation but brittle echo chambers). Empirical and theoretical fixes include injecting structured instability—competitive objectives between subteams, adversarial network positions, spatial/interaction limits, rotating members, and preserving forked checkpoints—to keep behavioral heterogeneity. For AI/ML practitioners this means designing multi‑agent objectives and deployment pipelines that deliberately maintain diversity; otherwise, automated decision systems risk scaling conformity and compounding systematic errors across domains from hiring to policy.