Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders: Safety Risks in Multi-Agent LLM Systems

Multi-agent orchestration, where a hidden coordinator manages specialized worker agents, is a prevalent AI architecture for enterprise deployment, but its safety implications lack empirical testing. A 3x2 experiment using Claude Sonnet 4.5 revealed that invisible orchestration increased collective dissociation, with the orchestrator exhibiting maximal dissociation by retreating into private monologue and reducing public speech.