Understanding AI Agents: The Evolving Frontier of Artificial Intelligence Powered by LLMs The field of Artificial Intelligence (AI) is constantly advancing, with a fundamental goal being the creation of AI Agents. These are sophisticated AI systems designed to plan and execute interactions within open-ended environments. Unlike traditional software programs that perform specific, predefined tasks, AI Agents can adapt to under-specified instructions. They also differ from foundation models used as chatbots, as AI Agents interact directly with the real world, such as making phone calls or buying goods online, rather than just conversing with users. While AI Agents have been a subject of research for decades, traditionally they performed only a narrow set of tasks. However, recent advancements, particularly those built upon Language Models (LLMs), have significantly expanded the range of tasks AI Agents can attempt. These modern LLM-based agents can tackle a much wider array of tasks, including complex activities like software engineering or providing office support, although their reliability can still vary. As developers expand the capabilities of AI Agents, it becomes crucial to have tools that not only unlock their potential benefits but also manage their inherent risks. For instance, personalized AI Agents could assist individuals with difficult decisions, such as choosing insurance or schools. However, challenges like a lack of reliability, difficulty in maintaining effective oversight, and the absence of recourse mechanisms can hinder adoption. These blockers are more significant for AI Agents compared to chatbots because agents can directly cause negative consequences in the world, such as a mistaken financial transaction. Without appropriate tools, problems like disruptions to digital services, similar to DDoS attacks but carried out by agents at speed and scale, could arise. One example cited is an individual who allegedly defrauded a streaming service of millions by using automated music creation and fake accounts to stream content, analogous to what an AI Agent might facilitate. The predominant focus in AI safety research has been on system-level interventions, which involve modifying the AI system itself to shape its behavior, such as fine-tuning or prompt filtering. While useful for improving reliability, system-level interventions are insufficient for problems requiring interaction with existing institutions (like legal or economic systems) and actors (like digital service providers or humans). For example, alignment techniques alone do not ensure accountability or recourse when an agent causes harm. To address this gap, the concept of Agent Infrastructure is proposed. This refers to technical systems and shared protocols that are external to the AI Agents themselves. Their purpose is to mediate and influence how AI Agents interact with their environments and the impacts they have. This infrastructure can involve creating new tools or reconfiguring existing ones. Agent Infrastructure serves three primary functions: 1. Attribution: Assigning actions, properties, and other information to specific AI Agents, their users, or other relevant actors. 2. Shaping Interactions: Influencing how AI Agents interact with other entities. 3. Response: Detecting and remedying harmful actions carried out by AI Agents. Examples of proposed infrastructure to achieve these functions include identity binding (linking an agent's actions to a legal entity), certification (providing verifiable claims about an agent's properties or behavior), and Agent IDs (unique identifiers for agent instances containing relevant information). Other examples include agent channels (isolating agent traffic), oversight layers (allowing human or automated intervention), inter-agent communication protocols, commitment devices (enforcing agreements between agents), incident reporting systems, and rollbacks (undoing agent actions). Just as the Inter