ixio
Code review · catching real bugs · updated 2026-05-20

Which agent reviews your code best?

7reviewers/122real bugs/50 · 5PRs · repos/3LLM judges
A code-review agent doesn't write code — it reads a pull request and flags the bugs before they ship. This ranks 7 of them on 122 real, merged bug-fixes replayed across 50 PRs from 5 production codebases, graded by a 3-judge LLM panel.

Reviewersranked by F1

source: Tenki benchmark ↗
#Reviewer
1
Tenkireview tool
68.9%29.9%
41.7
2
Devincoding agent
36.1%47.3%
40.9
3
Cursorcoding agent
32%51.3%
39.4
4
CodeRabbitreview tool
28.7%25%
26.7
5
Greptilereview tool
36.1%15.9%
22.1
6
Copilotreview tool
24.6%18.9%
21.4
7
Graphitereview tool
3.3%50%
6.2
How it's measured

Each PR contains a real, merged bug-fix from cal.com (TS), Sentry (Python), Grafana (Go), Keycloak (Java), Discourse (Ruby). The pre-fix diff is replayed into a clean fork and every tool reviews it at default settings — no custom rules, full repo context. A bug counts as caught only if a line-level comment pinpoints the faulty code and explains its impact, and ≥2 of 3 independent LLM judges agree.

Reviewers vs coding agents

Five are dedicated review tools (Tenki, CodeRabbit, Greptile, Copilot, Graphite); two are coding agents pointed at review (Devin, Cursor) — the same Cursor that shows up on the coding-agent board. The coding agents post fewer, higher-precision comments; the dedicated reviewers cast a wider net.

One source for now (the only public code-review benchmark). Snapshot verified 2026-05-20; it refreshes as the benchmark updates or a second, independent one appears.