This paper proposes a multi-agent collaborative system framework for Artificial Internet of Things systems. The framework integrates six modules: perception, planning, action, feedback, memory, and toolset, forming a closed-loop system. It introduces a dual-driven response mechanism combining user commands and IoT alerts, overcoming task response limitations. A hierarchical task instruction prompt architecture aids intent understanding and Chain of Action based reasoning, breaking complex tasks into actionable steps. Three “plug-and-play” actions knowledge augmentation, domain knowledge retrieval and workflow orchestration. A workflow-based language-to-code conversion method improves control script generation reliability. A closed-loop self-correction process ensures precise IoT system tool calls. Control scripts are locally stored and reused via LLM-based instruction recognition. Industrial case studies validate the framework's effectiveness.