this is one of the many strategies used in autogluon that enables it to outperform other autoML tools on most datasets:
https://arxiv.org/abs/2003.06505

https://arxiv.org/abs/2207.12560

One complaint people raise is regarding latency & complexity of deploying ensemble models, but there are many easy options to deal with this:
https://auto.gluon.ai/stable/tutorials/tabular_prediction/tabular-indepth.html#accelerating-inference