Agent golden paths: Building platform guardrails for AI that operates at population scale

When AI agents graduate from copilots to autonomous actors, most platforms break. They are designed for developers who read error messages, not systems making tens of millions of decisions daily without human intervention. The speaker draws on 19 years of building platforms where this failure mode is unacceptable, including digital health infrastructure serving millions of users. This talk shares lessons on agent-specific RBAC with blast radius limits, guardrails that prevent unsafe actions before execution, observability for non-deterministic systems, and validation layers that catch AI-generated configurations that look correct but fail at scale.