/TED: Alex Evers

Generative Adversarial PR Reviews

That's it. That's the whole post.

Well actually...

A practice my coworkers and I have been playing with:

  1. Generate spec/plan/tasks/code -> Create new PR
  2. Run preferred PR Reviewer Agent against PR
  3. Run preferred Agent against Generated PR Review
  4. Run preferred PR Reviewer Agent against Revised PR
  5. Run preferred Agent against Generated second pass PR Review
  6. ... this continues until the work week is exhausted.

Except, how much of the generated Review items become actual code changes?

I've noticed that I get a better sense of my own (generated) code, by having others point their Review-of-Shame Ralph Agents at it. By defending my code, improving what I think is necessary, saying 'wontfix' with increasing prejudice, I create a surface area of the PR that I stand by.

This seems to be a repeatable pattern that is resilient over time. Your preferred Agent/Reviewer agent may vary. Even running various ones against a branch before submitting a PR is useful, but after a point wearying.

My biggest complaint is that the careful consideration of code that was created is coming from the top down now, iteratively deepening, rather than the sideways and bottom-up that I usually take pre-LLM era.

Its exhausting because I have to read, think, and examine, retroactively. When normally I would read, think, design, build... once, then stand by my work and defend it with a reasonably open mind.

That last bit is largely lost here with LLMs. My work is not really my own until I get my hands into it, and thats now the last mile... not the first mile.