Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning