Agents-A1 achieves 1T-model performance through long-task training, not bigger parameters - TrendCloud