We propose AgentRecBench, the first large-scale and comprehensive benchmark that systematically evaluates agentic recommender systems and traditional recommendation methods across diverse scenarios.