3 Comments
User's avatar
Jenny Ouyang's avatar

Thanks for this article Shmulik! "Vibe merging" is the perfect name for something I've definitely been guilty of!

When you're reviewing AI-generated code and it passes CI/CD, it's so easy to assume the green check means the logic is sound. But CI/CD catches crashes and syntax errors, not intent.

I've started treating AI-generated PRs like they're from a junior dev who writes really fast but doesn't always understand the broader system. That mental shift forces me to slow down and actually review instead of just hitting approve because it looks plausible.

JP's avatar

The recall vs precision framing is the bit most people skip past. I went a different direction entirely. Instead of one agent trying to catch everything, I set up domain specialists (frontend, backend, DevOps) that only spin up when their files are touched. Way more focused output and you're not burning tokens on irrelevant checks. Wrote up the full pattern here: https://reading.sh/one-reviewer-three-lenses-building-a-multi-agent-code-review-system-with-opencode-21ceb28dde10?sk=805689479e932d817e8332104ec0a8c5

Shmulik Cohen's avatar

Very cool idea!

Could be implemented in the other code review systems by specialized rules, skills and tools probably. Code Review in general is not one size fits all, you need to fine tune it for your needs to get the most out of it