The rise of agentic AI has fueled excitement around agents that autonomously perform tasks, make recommendations, and execute complex workflows blending AI with traditional computing. However, creating such agents in real-world, product-driven environments presents challenges that go beyond the AI itself. Decoupling workflows, where agents, infrastructure, and other components interact fluidly without rigid dependencies, is crucial for achieving flexible, scalable integration. Event-driven architecture (EDA) powered by streams of events can help create a "central nervous system" for data, enabling seamless integration and flexibility as systems scale. In the context of PodPrep AI, an AI-powered research assistant, EDA is used to power an effective agentic system that processes data in real time, powering an AI-driven workflow without rigid dependencies. The use of Apache Flink and Confluent Cloud enables real-time RAG workflows, ensuring that question extraction works with the freshest available data. By decoupling components and using event streams, PodPrep AI demonstrates how EDA can enable real-world AI applications to scale and adapt smoothly.