How I Review AI Code - (Meta Senior Staff Engineer)

AI can write more code than any engineer can manually review. Here is the risk-based workflow I use to review AI generated code.

Chapters:
00:00 Should you still read AI-generated code?
01:30 The AI or Not demo
02:34 Code review is a gradient
03:47 Codebase trees and blast radius
05:05 Build behind feature gates
06:31 Require proof in every PR
09:00 Leaf code vs. trunk code
10:20 Safety over stylistic nits
12:31 Independent agent reviews
14:05 Codex on GitHub and PR babysitting
16:20 Merge-ready is not launch-ready
17:52 Experiments, canaries, and rollback
19:32 My current answer
20:04…

@jfrowiess I’m building a risk based classification system for our code reviews. Non risky stuff can be merged without human code reviews, more risky stuff will need a human or two reviewing them. 1 like

@JohnKimLife Yea I heard people are using JEV for this use case :) 1 like

@jfrowiess @JohnKimLife not yet myself but that’s the next step

@tam-mi-lv Yeah, code review is tricky part. I’m actually almost doing what you suggested in terms of code review. I personally still think that even though GPT models are very smart and capable they generate overengineer code, even Astra is doing this. So my preference would be still claude models for generating actual code, because it’s way much easier to read! 2 likes

Author posts

How I Review AI Code - (Meta Senior Staff Engineer)