Most organisations have spent years building an identity fabric that connects people, workloads and services through shared identity capabilities. Now AI agents are arriving, and many teams are responding by creating a separate stack for them: new tools, new credentials and new blind spots. The result is a parallel identity silo for the fastest-growing and least predictable identity type in the enterprise.
This session argues for a different approach: treating agents as a first-class identity type inside the fabric you already have. It walks through what actually needs to change, layer by layer. That includes how agents are registered, issued an identity and assigned a human owner; how delegation and on-behalf-of chains flow through token exchange; how authorization moves from static roles to runtime, task-scoped policy decisions at each tool call, including MCP-connected tools; and how governance and detection extend to agent behaviour, not just agent access. Drawing on real architecture work, the session shows which existing fabric components can be reused, where integration points such as API gateways and policy engines need to be extended, and what to build first.