Preprint: 'Infrastructure for AI Agents' by Seth Lazar, Elija Perrier and co-authors

In this paper, Alan Chan, Kevin Wei, Sihao Huang, Nitarshan Rajkumar, Elija Perrier, Seth Lazar, Gillian K. Hadfield, Markus Anderljung investigate the infrastructure we need to bring about the benefits and manage the risks of AI Agents. They identify three functions for agent infrastructure: 1) attributing actions, properties, and other information to specific agents, their users, or other actors; 2) shaping agents' interactions; and 3) detecting and remedying harmful actions from agents. They then propose infrastructure that could help achieve each function, explaining use cases, adoption, limitations, and open questions.

Read the full paper here.