Risks and Controls for Multi-Agent Systems: an analytical framework for deployment of AI agents across organisational boundaries

Multi-agent failures emerge from interactions themselves, not just individual agent errors. The framework distinguishes between risks that arise when AI agents interact with each other—which can be unpredictable and difficult to prevent—versus failures within single agents, requiring organizations to prepare for novel failure modes that don't exist in single-agent systems.

Organizational boundaries create uncontrollable risks. Once AI agent interactions cross an organization's perimeter (such as when agents interact with partners, customers, or unknown parties online), "no single organisation can fully see, control or govern them," meaning traditional internal governance and safety measures become insufficient.

Three governance tiers require different control strategies. The framework identifies singular governance (one organization controls all agents), federated governance (multiple organizations with agreed rules), and open environments (no central authority). Each tier presents distinct risks and requires different control approaches, with open environments presenting the most significant governance challenges.

Critical governance gaps exist where no single actor can act. The report identifies situations where necessary controls fall between organizational responsibilities—neither individual companies nor existing authorities are positioned to implement them—requiring new forms of collective action and coordination to address.

Policymakers and organizations need shared standards for open environments. The framework highlights that in open internet environments, "shared standards are adopted voluntarily if at all," suggesting the need for industry coordination and potentially new governance mechanisms to establish baseline safety practices across organizational boundaries.