I've thought that after releasing my current game, I could play around with large scale real time population simulation.

An interesting fact I learned is that there does not seem to be a significant performance bottleneck. Nagel published a paper in 1995 with an algorithm that could simulate the entire LA metro traffic on 1990s hardware. Actual real time large scale simulation was shown in the early 2000s etc.

AVX and hybrid GPU/CPU ECS can handle millions of "things". Factorio handles millions of belt objects and traffic is actually simpler because it can use cellular automata. Also peak traffic in a 4M city is only something like 150k-500k cars. HPA and caching could handle pathfinding at that scale.

Other parts of agent based simulation happen so rarely that they also should not take too much performance. Many things have to be updated at a minute scale, so may as well be millions of years for a CPU.

The translation in the post does not really make sense. At that point it's not agents anymore. Agents act based on their own ruleset, it's what makes them agents. But absolutely many things can be simulated at macro level that directly impact the agents behavior, because they simply change the simulation parameters.

The issue is mostly that realism is not always fun. If something realistic happens that the player does not understand it sucks. This is the core tension in these games I feel.

Agents will give probably the most realistic outcome, but not sure they make for the most fun one.

Reading this article made me realise that none of the people in my hometown played games like these. Ever. Or at least, not enough of them, because for the last thirty years, the city has been all about the income and the building, and the infrastructure can go suck some hairy balls...