Preprint: Infrastructure for AI Agents with Seth Lazar, Elija Perrier and Alan Chan
Autonomous AI agents are increasingly capable of navigating complex, open-ended environments and complete tasks such as browsing the web, making purchases, forming plans. As these systems proliferate, we will need more than just internal safeguards to ensure they behave safely and predictably.
In this preprint, MINT researchers Seth Lazar, Elija Perrier and Alan Chan, among a broader research team, introduce the concept of agent infrastructure – shared, external systems and protocols that shape how agents interact with the world and each other. Just as the internet depends on protocols like HTTPS, a mature AI ecosystem will depend on infrastructures that mediate agent action.
The paper outlines three key functions of such infrastructure:
Attribution of actions and responsibilities to specific agents or humans
Interaction protocols for safe and effective agent communication
Remediation systems for detecting and addressing harmful behaviour
By mapping emerging challenges and research directions, this work contributes to the growing field of sociotechnical AI safety, where systems aren’t just engineered, but embedded in institutional and social contexts that help them serve human values.
Read the full paper here.