AI agents are no longer experimental — they're running in enterprise production environments, accessing sensitive systems, handling regulated data, and executing actions on behalf of employees. And yet most organizations cannot answer a basic question: when an AI agent does something it shouldn't, who is accountable? According to the State of Autonomous Agent Governance, 49% of security executives have discovered an AI agent operating with permissions it should not have had — and 61% found out after the fact.

The challenge isn't negligence. It's architectural. Identity controls built for human accounts — quarterly access reviews, session-level authentication, perimeter-based policy — were never designed for agents that act autonomously, at millisecond cadence, across dozens of connected systems. As enterprises move from dozens to thousands of AI agents, the gap between human governance models and agent realities widens fast. The same research found that only 37% of organizations are actively limiting agent access and the rest are operating on trust.

This session examines how leading security teams are rethinking the accountability model for AI agents: what it means to tie every agent action to a named human owner, how governance can operate inline with agent workflows rather than periodically around them, and what "earning autonomy one rung at a time" looks like in practice — frameworks that let organizations extend agent responsibility as trust is established, without sacrificing auditability or control.

Key Takeaways:

1. A framework for deciding what agents need governed access to versus what should be blocked entirely

2. How identity-native governance differs structurally from perimeter-based AI security controls

3. Practical approaches to building human accountability layers over autonomous agent workflows

4. Criteria for when automated governance decisions can be trusted — and when a human must stay in the loop