ByteDance Seed released EdgeBench, a long-term benchmark measuring AI agents' real-world continuous learning. It covers 134 tasks across six fields, each requiring at least 12 hours of work, based on 38,000 hours of interaction data. Agents show a log-sigmoid learning curve with R²=0.998; cutting-edge models double learning speed every three months. 51 tasks and the full evaluation framework are open-sourced.