AI agents now authenticate to cloud services, query protected data and invoke tools on behalf of employees, contractors and other systems. After an agent inherits broad credentials or loses the user’s authorization context, accountability can disappear across a chain of legitimate-looking API calls. In this roundtable, we examine AI agents as a distinct identity and behavior problem.

Participants will compare approaches for assigning agent identities, limiting permissions by task, carrying user authorization through a workflow, and detecting activity that falls outside the agent’s intended purpose. We will also cover behavioral baselines, shared memory, tool access, emergency shutdown, and the audit records needed across agency and contractor environments.

We will work through a deceptively simple test: after an agent takes action, can your team prove who requested it, what authority it used and whether its behavior remained within policy?