From experience: A static type system is MUCH better for agents than dynamic. I've done a few projects both ways - the agent is just much, much better with the static system because it sees compile errors right away, and also has access to better static analysis tools. I don't know how much the types themselves help the LLM, it's better context but more tokens, but the agent seeing errors is just a huge difference.
It's also much easier for you as a developer to digest the code if you need to refactor because you got lazy and accepted some slop.