The rapid adoption of AI agents across diverse applications highlights the critical need for robust evaluation frameworks. Drawing on hands-on experience building over 30 agents, this talk explores the principles of AI agent evaluation that balance rigor, practicality, and adaptability. Traditional machine learning and recommendation system metrics—such as precision, recall, and ranking quality—remain essential benchmarks, offering a foundation for measuring agent performance. Yet, in agent-driven systems, content generation quality emerges as the top priority, requiring human-centered measures of coherence, creativity, and utility. A central challenge lies in generating and curating a golden dataset that is both representative and scalable, enabling reproducible and meaningful evaluation. Finally, we present strategies for building iterative evaluation loops, including the design of “evaluation agents” that autonomously test, critique, and refine AI outputs. Together, these principles offer a roadmap for practitioners to build trustworthy, high-performing AI agents that continuously improve in real-world settings.