Current mainstream AI Agent architectures (ReAct, Chain-of-Thought, AutoGPT) are designed with a single-user perspective by default, and face fundamental challenges in multi-user collaborative scenarios, including ambiguous context attribution, intent conflicts, and difficulty in state synchronization.