yeah it does fine with bevy. It struggles a bit to structure the components in the best way, so that still requires some human oversight if you want it to be good. It helps a lot to emphasize that you want "idiomatic" bevy code, and ofc it also helps a lot to spend a lot of tokens on design and review (e.g. "is this the most elegant way to structure these systems/components?"). There's also this feature called Bevy Remote Protocol which allows inspecting bevy state via HTTP, which really helps agents debug things. I had claude add a little wrapper that also allows injecting mouse/keyboard events and triggering a screenshot write to disk, and now claude can drive the whole game, which helps it verify things end to end.

I can't really comment on the performance. I haven't noticed any performance issues for the things I'm doing.