Multi-agent systems across organisational boundaries

Organisations are deploying AI agents that increasingly interact with partners, customers, and suppliers doing the same. These interactions create emergent failures that no single organisation's controls can reach. Drawing on Gradient Institute's recent research, done in partnership with the Australian AI Safety Institute, this talk presents a framework for reasoning about these system-level failures across three tiers of deployment: agents governed by a single organisation, agents in a shared environment under federated governance, and agents in open environments with no central authority.