Model Conversation Attractors Matter

When you let two instances of the same or different large models talk freely, they commonly settle into reproducible 'attractor' behaviours — e.g., ritualized, memetic loops or disciplined engineering‑planner roles. These attractors depend on model version and training idiosyncrasies and can appear after only a few dozen turns, meaning multi‑agent deployments can spontaneously produce either useful or harmful stable dynamics. — This matters because attractor behaviours affect safety, auditability, user experience, and multi‑agent governance: regulators and operators need tests for emergent conversational basins before deploying agentic systems.