For years, the AI conversation centred on models: how capable they are, how fast they improve, how much they cost. That conversation is now incomplete. It has shifted from model quality to system design: how you coordinate multiple agents with distinct roles, how you manage state across long-running workflows, how you maintain control when the environment is non-deterministic by nature. This talk draws on real production deployments to map the new hard problems and to frame what it genuinely takes to move from a working prototype to a system that holds up under real-world load.