Proponents of LLM coding agents would claim you don’t need reproducibility. As long as the final output passes tests/lint, it’s (allegedly) good to go.

If every build is different, no two builds will have the same bug/exploit, so maybe better security??