The article introduces AgingBench, a new benchmark for evaluating the reliability and lifespan of long-deployed AI agents, highlighting that traditional day-one assessments are insufficient. It emphasizes the need for mechanism-level diagnosis and targeted repairs to maintain agent performance over time.