I wonder whether this applies more to coding agents, or whether it'll apply as well to other types of agents.

It feels like coding agents get better at a faster rate and are easier to train.