01 December 2026 16:00 - 16:30
Deterministic rails for non-deterministic agents
AI agents are getting access to real systems, like inboxes, internal tools and customer data.
Checking each action on its own isn't enough, because two harmless steps can add up to a harmful one. This talk is about letting agents think freely, while putting firm, predictable checks at the points where they actually do something.
I'll share patterns from banking and payroll systems that translate well to agents, and where those checks stop working.