This paper introduces ICR, a framework to evaluate communication in LLM multi-agent systems. It helps disentangle the effects of communication, architecture, and reasoning on system performance.