A new episode of the Practical AI podcast explores the critical gap between experimental AI agents and production-ready systems, with hosts Hamza Tahir and Daniel Whitenack diving into how machine learning operations (MLOps) principles are reshaping agent development at scale. The conversation covers essential infrastructure components—including workflows, agent harnesses, and fleet management—that distinguish reliable production agents from proof-of-concept prototypes. As organizations attempt to deploy agents beyond controlled environments, observability, replayability, and resilience have emerged as non-negotiable requirements. ZenML's new open-source project, Kitaru, exemplifies this infrastructure shift. Designed to help developers build resilient, replayable, and observable agent systems, Kitaru addresses pain points that teams encounter when transitioning from research to production. The episode highlights how the ML tooling landscape has expanded dramatically, with over 84 new tools now available to support production AI challenges, signaling growing recognition within the industry that sustainable agent deployments require operational maturity, not just model capability. The discussion underscores a broader maturation trend: as generative AI reaches production scale, organizations are learning that managing agent complexity, auditability, and reliability demands the same rigor that MLOps has brought to traditional machine learning workflows. This shift from "demo to durable" represents a fundamental transition in how teams approach AI system development and deployment.