imo, the real value of Ada/SPARK today is that it enforces a clear split between specification and implementation, which is exactly what your LLM needs.

You define the interface, types, pre/post conditions you want in .ads file, then let the agent loose writing the .adb body file. The language’s focus on readability means your agent has no problem reading and cross referencing specs. The compiler and proof tools verify the body implements the spec.