What Breaks When Agentic AI Reaches Production?

Cockroach Labs VP Quentin Packard identifies six critical infrastructure failures that surface when agentic AI moves from proof-of-concept to production: memory state management, thundering herd problems, agent identity, blast radius containment, observability gaps, and unit economics. Rather than blaming AI model limitations, the article frames agentic AI as forcing enterprises to solve classic distributed systems problems at a new scale. The piece emphasizes that durable state management, autonomous actor governance, and databases designed for high-throughput writes with strong consistency are prerequisites for reliable agentic deployments.