📄 中文摘要
多智能体系统日益普及，CrewAI团队、LangChain代理团队和MCP连接助手等协同工作。然而，缺乏治理机制使得每个代理的行为难以控制，可能导致被攻击或不当行为的代理访问不应接触的数据。为了解决这一问题，提出了基于政策的访问控制方法，以确保每个代理的权限得到合理管理，从而防止潜在的数据泄露和安全隐患。此方法强调了制定明确的政策、进行必要的审批以及建立审计轨迹的重要性，以实现对多智能体系统的有效治理。
多智能体人工智能系统的治理：政策、批准与审计轨迹
出处: Governing Multi-Agent AI Systems: Policies, Approvals, and Audit Trails
发布: 2026年3月21日
📄 English Summary
Governing Multi-Agent AI Systems: Policies, Approvals, and Audit Trails
Multi-agent systems are increasingly prevalent, with CrewAI crews, LangChain agent teams, and MCP-connected assistants working collaboratively. However, the lack of governance mechanisms makes it difficult to control the actions of each agent, potentially allowing compromised or misbehaving agents to access sensitive data. To address this issue, a policy-based access control approach is proposed to ensure that the permissions of each agent are managed appropriately, thereby preventing potential data breaches and security risks. This approach emphasizes the importance of establishing clear policies, necessary approvals, and audit trails for effective governance of multi-agent systems.