Sometimes I think that if we are willing to burn tokens and rely on compilers in a loop we should be using languages that can catch as many errors as possible at compile time.
Like Ada or Ocaml or Haskell or something.
Or even require like MC / DC testing or MISRA C verification.
Or even some of the languages that apply Hoare checks, like Ada SPARK or FRAMA-C or VALE or whatever.
I don't have the budget to test how that would work but it seems interesting.
I presume haskell applications similar to claude code are well represented in the training data.
There was a post recently from a guy who used Haskell with AI agents. If I remember correctly he said the type checking did catch errors. The problem was the compiler was very slow, which slowed down the feedback loop and output. I think he switched to Python.