Feature quality is often a deciding factor in predictive models’ performance. However, effective features require crafting with domain expertise, which is difficult to scale across new applications. Existing feature engineering approaches still fall short in generating domain-aligned, high-quality features. We present Eureka, a generalizable LLM-driven agentic framework that automates feature engineering with evolving intelligence. Our approach has three main components: an Expert Agent that encodes domain knowledge to evaluate feature quality, an Automated Feature Generator that translates feature designs into executable code, and a RL (Reinforcement Learning) Feedback Loop that connects the two components and enables continuous learning. Evaluated on 7 public benchmarks spanning healthcare, finance, and social domains, Eureka consistently achieves superior performance. Furthermore, to validate effectiveness in real-world deployment, we apply Eureka end-to-end to cloud GPU resource demand prediction at Alibaba Cloud, a particularly challenging task due to sparse historical data and rapidly evolving workloads. At scale, Eureka improves demand fulfillment rate by 16% and reduces computing resource migration rates by 33%.