The path from AI agent prototype to reliable production system requires much more than a well-trained model, according to industry experts discussing the state of agent infrastructure. In a recent conversation, developers and researchers explored the growing recognition that machine learning operations (MLOps) principles—traditionally applied to model deployment—are essential for building durable, scalable agent systems. The discussion highlighted challenges specific to generative AI agents, including workflow complexity, system reliability, and the difficulty of observing and debugging agent behavior in production environments. ZenML, a machine learning orchestration platform, is addressing these challenges with Kitaru, a new project designed to help developers build resilient, replayable, and observable agent systems. The platform aims to bridge the gap between successful demos and production-ready deployments by providing infrastructure for managing agent fleets, handling edge cases, and maintaining system transparency. As enterprises increasingly invest in AI agents for business-critical functions, the need for robust MLOps tooling has become urgent, suggesting that infrastructure and operations will play as crucial a role in agent success as the underlying models themselves.