Towards Efficient Llm Grounding For Embodied Multi Agent Collaboration