A Meta infrastructure tech lead argues that as AI agents move into production, the core challenge shifts from model capability to infrastructure reliability. Traditional cloud infrastructure was built for deterministic, short-lived requests, but autonomous agents are stateful, long-running, and probabilistic — creating a fundamental mismatch. Key failure modes include retry amplification, recursive reasoning loops, and memory consistency issues rather than hallucinations. The recommended approach treats AI agents as distributed systems: models should only generate proposals while a policy engine and execution gateway enforce decisions. An 'agentic control plane' analogous to Kubernetes is needed for scheduling, memory coordination, policy enforcement, and workload routing. Existing distributed systems patterns — circuit breakers, rate limits, resource quotas — can be adapted for agentic reliability. The competitive advantage in AI is shifting from better prompts and models to more reliable infrastructure.