AI Review Arena
Make AI models argue with each other before your code ships.
English | 한국어
The Idea
You ask one AI to review your code. It finds 12 issues. But which ones are real?
Arena's answer: make three AIs fight about it.
┌──────────────────────────────────────────────────────────┐
│ SINGLE AI REVIEW │
│ │
│ You ───► One AI ───► "12 issues found" │
│ │
│ But... which are real? You have to check all 12. │
└──────────────────────────────────────────────────────────┘
vs.
┌──────────────────────────────────────────────────────────┐
│ ARENA REVIEW │
│ │
│ You ───► Claude ───┐ │
│ Codex ───┼──► They argue ───► 5 real │
│ Gemini ───┘ with each other issues │
│ │
│ 3 AIs independently review, then cross-examine │
│ each other's findings. Fake issues get eliminated. │
│ Real issues get confirmed with higher confidence. │
└──────────────────────────────────────────────────────────┘
How The Fight Works
Three AI families review your code separately, then challenge each other in 3 rounds:
ROUND 1 ROUND 2 ROUND 3
Independent Review Cross-Examination Defense
────────────────── ───────────────── ───────
Claude: "I found Codex: "Claude's Claude: "No, look
a SQL injection finding #3 is a at line 42 — user
at line 42" false positive, input goes directly
this input is into the query
Codex: "I found already sanitized" without escaping.
a race condition Here's proof..."
at line 89" Gemini: "Actually,
I agree with ───► CONFIRMED
Gemini: "I found Claude — the confidence: 92%
unused imports sanitization
at line 7" misses Unicode" ───► DISMISSED
(false positive)
What survives this fight = what you should actually fix.
But Wait — It Does Way More Than Code Review
Arena isn't just a reviewer. It's a full lifecycle system that handles everything from "I have an idea" to "ship it."
"Build an OAuth login"
│
▼
┌─────────────────────────────────┐
│ ARENA PIPELINE │
│ │
│ 1. Analyze your codebase │ ← learns your coding style
│ 2. Research best practices │ ← searches the web
│ 3. Check compliance rules │ ← platform guidelines
│ 4. Debate implementation │ ← AIs argue about HOW to build it
│ 5. Build it │
│ 6. Review with 3 AI teams │ ← the fight described above
│ 7. Auto-fix safe issues │ ← fixes trivial things automatically
│ 8. Generate tests │ ← writes regression tests
│ 9. Final report │ ← pass/fail verification
│ │
└─────────────────────────────────┘
And it works for three domains, not just code:
| Code | Business | Documentation |
|---|
| Routes | A-F | G-I | J-K |
| Example | "Build OAuth" | "Write pitch deck" | "Review API docs" |
| Reviewers | 12 specialized agents | 10 specialized agents | 6 specialized agents |
| Special | Threat modeling, static analysis | Red team, quant validation | Code-doc drift detection |
It Turns On Automatically
You don't call Arena. Arena calls itself.
Every request you make to Claude Code gets routed through Arena automatically:
You say: Arena does:
───────────────────────────── ─────────────────────────────
"Build a login page" → Route A: Full lifecycle
"Fix this typo" → Route F: Quick fix (instant)
"Review this PR" → Route D: Multi-AI review
"Write a pitch deck" → Route G: Business pipeline
"Are the docs accurate?" → Route J: Doc review pipeline
"Refactor this module" → Route E: Refactoring pipeline
"Research auth best practices"→ Route B: Deep research
HajinJ