AI Code Review: 30K Bugs Lighter, 50% faster
AI Code Review: 30K Bugs Lighter, 50% faster
Last month, we launched AI Code Review, our developer tool that automatically catches bugs, finds performance issues, and helps you ship PRs faster.
30 days later, here's what’s new:
Testing AI Code Review on our own code
So far, AI Code Review has caught more than 30,000 bugs. A couple of them were in Sentry Staff Engineer Ryan Brooks’ side project, an app that builds and maintains training plans for endurance athletes. Think of it like a virtual coach that adapts your workouts based on your Strava performance.
It caught a scheduling logic bug:
“Part of my app generates training ‘phases’ working backwards from a race date. I made some tweaks to this algorithm, and Sentry correctly caught a bug where two phases could potentially overlap — a bug that definitely shouldn’t be possible. I copied the comment from my PR into Claude Code and the bug was fixed.”
And a mistake that would’ve definitely impacted users:
“I also added some logic to my onboarding flow that prevents users from adding a race less than 3 weeks away — because, let’s be honest, you’re not going to run a marathon next weekend. But I use the same React component to edit races, and Sentry caught a bug that would’ve blocked users from editing races that were already scheduled. Huge catch.”
Faster comments with copy-and-paste prompts:
50% better performance
Our reviews were taking longer than anyone liked, so we reworked the review pipeline to bring average latency down by roughly 50%.
Here’s how:
For some tasks that required less reasoning, we switched to more performant models, while tweaking the prompt to maintain output consistency.
We applied a “thinking budget” for all steps, and maximum iterations to the hypothesis and verification steps, to prevent overthinking.
We ran evals & spot checks to ensure the performance improvements didn’t negatively impact the reviews.
In short, reviews are now landing faster and leaner, and we’re continuing to tune that balance.
Clear, actionable PR comments
We’ve revamped how comments are structured to improve readability, and added an actionable AI prompt. Here’s the new structure:
Detailed Analysis – what’s wrong and why
Suggested Fix – how to fix it
New: AI Prompt – a copy and pasteable prompt to send to your favorite AI agent for an instant patch
A Claude Skill that automates bug fixes
Cody De Arkland is literally always tinkering. To help him move faster, he created a Claude Skill that lets you skip the copy and past step, connecting Claude Code directly to AI Code Review outputs. It takes flagged issues, validates them against your repo context, and auto-generates verified patches. Give it a try and let him know what you think.
ICYMI: The AI Code Review workshop
We also hosted a live workshop where we walked through setting up and using AI code review. If you missed it (or want a replay), you can watch it on demand here.
Try it (and tell us what you think)
If you haven’t tried AI Code Review yet, get started here.
If you have tried it, share your feedback.
