Autonomous robots often face challenges in robust perception, semantic understanding, and safe interaction when performing manipulation tasks in complex unstructured environments. To address this issue, we propose a novel Item-Level Multi-Agent (ILMA) interaction framework. ILMA enables robots to dynamically identify task-relevant objects and represent them as virtual communicating agents coordinated by a central Multi-modal Large Language Model (MLLM). These agents proactively provide contextual information and real-time feedback through a lightweight structured messaging protocol, deeply participating in task execution. To support the training and evaluation of ILMA, we introduce HomeTask-101, a new benchmark containing 101 diverse household tasks richly annotated for the ILMA paradigm. Extensive experiments demonstrate ILMA’s superior performance: on a long-horizon task benchmark, ILMA improves the average task completion length by 18.6%; in 50 manipulation tasks from MetaWorld, the average success rate is 10.8% points higher than baselines; and on complex unseen tasks with a real-world robot from HomeTask-101, it surpasses the previous state-of-the-art methods by over 30%. These results strongly demonstrate the effectiveness of ILMA in significantly enhancing task success rates, execution efficiency, and robustness.