The Evolution Stable State of Predictive Processing in a Minimal Environment

People are starting to try and simulate this kind of thing by simulating lots of little artificial active inference agents all doing their predictive processing together in a minimal context. What tends to happen is something very interesting you basically get a partition into an into two groups and it's usually fifty-fifty so why don't they just merge? I think the answer lies in the initial conditions and the heterogeneity that people people bring to the table when they start their exchanges but also something more subtle on top of that which is something that comes when you start to think about agents that plan.