Many organizations invest heavily in monitoring and observability for AI agents, assuming that comprehensive visibility will improve safety. In practice, detailed logs often provide an excellent record of failures without preventing them.

This session challenges the idea that observing agent behaviour is the same as controlling it. Using examples from production agent deployments, we will examine situations where agents gained unintended access, where credentials crossed boundaries they should not have crossed, and where interactions between systems produced outcomes that were never anticipated. In each case, the focus is not on detecting failures after they occur, but on understanding how architectural controls could have prevented them.

The talk explores governance as an architectural discipline built around enclosure, entitlement, and explicit permissions. Topics include credential scoping, integration-level access control, capability-based permissions, data boundary enforcement, and designing agents with only the access required for a specific task. Attendees will see how these patterns differ from observability-focused approaches and why prevention requires more than monitoring.

The session concludes with a practical framework for treating agent safety as a property of system design rather than an operational afterthought. The goal is to help teams recognise where monitoring is being used as a substitute for governance and understand the architectural controls required to prevent common failure modes before they occur.

What You Will Learn:

• Why observability and monitoring alone cannot prevent many common agent failure modes
• How credential scoping, capability-based access control, and data boundary enforcement can reduce risk by design
• A framework for approaching agent governance as an architectural concern rather than a monitoring problem

Who Should Attend:

• Platform engineers
• Security engineers
• Enterprise architects
• AI engineers building agent-based systems
• Staff and principal engineers
• Technical leads
• Engineering managers responsible for AI adoption
• Teams designing agent platforms and integrations