> you'll get the best results with agents from the same tribe, others will trip over the mountain of idiosyncratic metaphors.

That's actually not what people found in practice. There's some research from the folks making smol-agent that you can get better results by randomly alternating calls between gpt and opus. The overall task solving rate is better than either one of them. So ymmv depending on task use (this was for coding).