北京大学：2026年智能体伦理治理研究报告摘要版（英文版）

Al agents are built on large models, but their defining distinction liesnot in stronger text generation, but in the capacity to pursue goals throughsustained action. They can plan multistep tasks, invoke tools, and querydatabases, thereby forming a coherent chain of execution. System outputis no longer the endpoint; it becomes an input for subsequent decisionsand actions. Mainstream Al agents typically comprise a foundation model,planning, memory, tool, and feedback modules, which respectivel