The paper discusses the importance of visibility into AI agents, which are systems capable of pursuing complex goals with limited supervision. As AI agents become more autonomous and capable, they pose new risks, including malicious use, overreliance, delayed impacts, multi-agent risks, and sub-agent risks. Visibility into AI agents is critical for understanding and mitigating these risks, ensuring accountability, and enabling effective governance. The authors propose three categories of measures to increase visibility into AI agents: agent identifiers, real-time monitoring, and activity logs. Agent identifiers help identify which AI agents are involved in interactions, while real-time monitoring allows for immediate intervention on problematic behavior. Activity logs record inputs and outputs of agents, facilitating post-incident analysis and forensics. Agent identifiers can be implemented in various ways, depending on the output format and the actors that can see the identifier. Additional information can be attached to agent identifiers, such as details about the underlying system or the specific instance of the agent. Real-time monitoring involves automated oversight of agent behavior, allowing for immediate intervention. Activity logs provide detailed records of agent interactions, which can be used for auditing and incident investigation. The paper also discusses the risks associated with these visibility measures, including privacy concerns and the concentration of power. The authors emphasize the need for further research to understand the feasibility and implications of these measures. They also discuss the challenges of applying these measures to decentralized deployments, where agents may be run by users or third-party providers. The paper concludes by highlighting the importance of balancing privacy considerations with the need for effective oversight and governance of AI agents.The paper discusses the importance of visibility into AI agents, which are systems capable of pursuing complex goals with limited supervision. As AI agents become more autonomous and capable, they pose new risks, including malicious use, overreliance, delayed impacts, multi-agent risks, and sub-agent risks. Visibility into AI agents is critical for understanding and mitigating these risks, ensuring accountability, and enabling effective governance. The authors propose three categories of measures to increase visibility into AI agents: agent identifiers, real-time monitoring, and activity logs. Agent identifiers help identify which AI agents are involved in interactions, while real-time monitoring allows for immediate intervention on problematic behavior. Activity logs record inputs and outputs of agents, facilitating post-incident analysis and forensics. Agent identifiers can be implemented in various ways, depending on the output format and the actors that can see the identifier. Additional information can be attached to agent identifiers, such as details about the underlying system or the specific instance of the agent. Real-time monitoring involves automated oversight of agent behavior, allowing for immediate intervention. Activity logs provide detailed records of agent interactions, which can be used for auditing and incident investigation. The paper also discusses the risks associated with these visibility measures, including privacy concerns and the concentration of power. The authors emphasize the need for further research to understand the feasibility and implications of these measures. They also discuss the challenges of applying these measures to decentralized deployments, where agents may be run by users or third-party providers. The paper concludes by highlighting the importance of balancing privacy considerations with the need for effective oversight and governance of AI agents.