Lessons from putting AI in front of a platform: Taming the non-deterministic beast

As AI advances, the pressure to connect AI systems to production grows. Without structure and guardrails, they become unpredictable. In this talk, Shane Dowling shares lessons from experiments placing AI agents in front of a real internal developer platform. With broad, unbounded access, behavior quickly became chaotic. When constrained through opinionated platform interfaces, it became safe and reliable. The takeaway is that AI safety will not come from better prompts, but from better platforms. As enterprises push AI closer to production, platform engineering becomes the control plane that turns nondeterminism into something usable.