An agent score replaced my review
One day to build, one day waiting for review — usually mine. Now PRs scored 4/5 or 5/5 by an agent ship without a human reading them. Months in production, no surprises.
Julio, can you review my PR? One of our biggest bottlenecks at Riqra was code review. A feature took a day to implement — and then another day waiting for review. Too often, the reviewer was me.
I didn't want to be the bottleneck again. But I also wasn't going to lower the bar on what reaches production.
What review was actually for
When I looked closely, my review wasn't really line-by-line reading. It was a trust check: does this fit the architecture, does it handle the risky cases, would I bet production on it? The real question was whether that trust check needed to be me.
We brought in an agent reviewer — Greptile. Every PR gets reviewed: feedback, risks, and a score from 1 to 5.
The rule that changed the math
Then the internal rule: PRs scored 4/5 or 5/5 can go to production without human review. If the engineer feels a PR needs human eyes, they ask for them. And responsibility stays exactly where it was — with the team, because they're the ones on call.
The score didn't replace judgment. It replaced waiting for mine.
We've been running like this for months, without production problems. The queue with my name on it is gone.
The part that isn't about the tool
The tool took an afternoon to set up. The real work was deciding what our bar for production actually is, writing it down, and moving accountability to the people shipping. An agent reviewer without that is just a second opinion nobody owns.
Founder question: is your review queue protecting quality — or protecting a person's place in the loop?