MES-replacement

Is there any hard evidence that using AI agents significantly improve the productivity in software development, excluding website development? I have been developing C(++) software for more than 30 years and I am still discovering things about the language that I had not fully grasped before, especially now that I am developing a C-compiler [1] myself. What I want to say is that developing quality software, a deep understanding of programming languages, is required. Good software engineering is not about produce lots of code, but keeping you code base lean/agile, avoiding technical depth, such that you velocity does not decrease over time. I am not convinced that that can be achieved with current day AI agents that only have a shallow understanding of software and that I doubt are able to have a deep understanding of large code base. So,

I am currently implementing a stack based language as an intermediate language for a C compiler that I am developing. I have developed a relatively straight forward compiler for the stack language to the M1 assembler from live-bootstrap. This stack language has local variables. It actually makes use of two stacks: One in the traditional sense and one for the local variables (and the return address for calls). It reads as a kind of very primitive post-fix C kind of language. See: https://github.com/FransFaase/MES-replacement and https://www.iwriteiam.nl/D2506.html#2b