A pioneer of machine teaching who led efforts at Microsoft argued that building reliable autonomous AI requires structured practice environments and multi-agent orchestration, not single-model scale-ups. The author said enterprise deployments stall because agents lack repeated, realistic experience and role-defined responsibilities akin to sports teams. The lessons have direct implications for university AI labs and professional programs: research that treats agents as teams will demand new curricula, simulation infrastructure and partnerships between academia and industry to validate agent behavior before real-world deployment.