Unconstrained self-organizing LLM teams fail to achieve synergy, according to Apple Machine Learning Research. The study finds that agents interacting freely without fixed workflows often perform worse than a single expert model. This suggests that emergent coordination is currently inefficient.