The integration of Large Language Models (LLMs) with Multi-Agent Systems offers a transformative approach to embodied intelligence, but faces significant challenges in partially observable environments and close-loop feedback. This paper presents a multi-stage framework that utilizes LLMs for coordinated multi-agent cooperation under partial observability. It decomposes complex tasks into three stages: Planing, where high-level instructions are translated into mid-level subtasks; Execution, where subtasks are converted into low-level actions; and Feedback, which involves real-time monitoring and adaptive re-planning. A closed-loop mechanism refines actions dynamically through environmental feedback and self-verification, enhancing robustness in dynamic settings. Validated in AI2Thor-simulated household tasks, the framework demonstrates effective multi-agent coordination and promising results.