Interesting experiment!

However, I think it lacks the most interesting information - which is a latency in tokens/sec and how it decays with increasing of parallelism. I also not exactly got what it has to do with AI Agents in particular.