Discussions about AI autonomy often arrive at the same question: trust. In many organisations, trust in agentic systems is based either on confidence gained from demonstrations or on policies that require human review. Neither approach is sufficient on its own once systems reach production.

This session argues that trust is an engineering property that should be designed into a system rather than assumed. It explores how trust can be earned gradually, verified through engineering controls, and revoked when conditions change.

Drawing on patterns from production agent systems, the talk examines how trust can progress through stages ranging from declared intent to agent-backed review and deterministic enforcement. It also explores classifying permissions according to reversibility and blast radius, applying verification at different stages of the development lifecycle, and maintaining trusted systems over time through ongoing review and cleanup.

The session focuses on how engineering rigor shifts as agentic systems become more capable. Rather than disappearing, it moves from writing software directly to designing the systems that generate, verify, and constrain it. Attendees will leave with a practical framework for treating trust as a property that can be deliberately engineered into agent-based systems.

What You Will Learn:

• How trust can be designed as an engineering property rather than treated as a policy or assumption
• Approaches to structuring permissions and verification based on reversibility, blast radius, and different stages of system operation
• Why trusted systems require ongoing maintenance to remain trustworthy over time

Who Should Attend:

• AI engineers
• Platform engineers
• Software architects
• Staff and principal engineers
• Technical leads
• Engineering managers responsible for AI systems
• Teams designing or operating agent-based platforms