When adding more agents to distributed training reduces total computation time proportionally, achieving near-perfect parallelization.