Deterministic AI code workflows when?

A lot of software engineering teams are moving away from full code reviews because they don’t have time, and some projects no longer accept pull requests because they don’t have time and they would prefer to have the prompt that created the PR.

This mirrors the same move we have seen many other times. We move from focus on the surface-level code to focus on the source, which somehow generates the surface-level code:

• When we got compilers we stopped reviewing the machine code and focussed on the high-level code.

• When we have a codebase in (say) C, and then decide to rewrite parts in a language like Nim which transpiles to C, we don’t really review the generated C. We focus on the Nim source. We don’t even commit the generated C to version control, sometimes.

The difference is that these generative processes are deterministic, fast, and automatic. The generation of surface-level code from prompts and specs is non-deterministic, slow, and interactive. So, for now we may still decide to commit the prompts and specs to version control, but we definitely commit the surface-level code.

For now! But the trend suggested by history and by recent practice is that we will move towards putting the specs and prompts into version control, and reviewing them carefully, and not committing the surface-level prompt. For this to work, we’ll need fast, deterministic, automated LLM workflows.