In fact, looking at the blog post, the agent orchestrating 16 GPUs is half as efficient as the agent using 1 GPU in GPU-time. Since it uses 16 GPUs to reach the same result as 1 GPU in 1/8 of the time.