Crystal Review
A multi-agent AI code reviewer that runs as a GitHub App on the company's repositories.
Key result
Engineers accepted 76 of 83 labeled comments (91.6%)
- Year
- 2026
- Role
- Solo
- Context
- Bielcrystal
Recorded on a local copy with demo data.
How it works
- 01
A planner splits each pull request into review tasks. Parallel reviewer agents inspect them, and a judge agent filters and ranks the findings.
- 02
Findings that can be checked are verified by executing code before they are posted.
- 03
On 26 production pull requests it left 89 inline comments. Engineers accepted 76 of the 83 labeled ones.
- 04
Rejection rate was 8.4%, against 11.7% for CodeRabbit and 31.2% for Kodus under the same labeling.
Screens
