As AI systems evolve from standalone models to increasingly autonomous agents, evaluation must evolve with them. Traditional benchmarks often focus on model-level accuracy, while real-world performance also depends on tools, memory, environments, context selection, orchestration, and the surrounding agent harness. This talk examines the shift toward system-level evaluation of LLMs and agentic AI, including robustness, safety, fairness, uncertainty, computational budgets, and reproducibility. It also considers multimodal agents operating on potentially manipulated evidence, where authenticity and provenance become part of evaluation. Drawing on our work in agentic evaluation and deepfakes, we outline practical principles for assessing AI systems in realistic settings and deployments.