What Do You Do While AI Codes? I Make Mine Argue With Itself.

Developer Debashish Ghosal explores the pitfalls of using AI for code review, arguing that most multi-model systems are merely performing 'theater' by anchoring subsequent models to the first model's output. To solve this, Ghosal developed 'AdversarialDebate,' a system that forces two LLMs to analyze code independently without shared context before engaging in a structured debate. By preventing models from seeing each other's verdicts until they have committed their own, the system successfully eliminates bias and consensus-seeking behavior. Testing across 150 artifacts showed an 88.7% accuracy rate in identifying issues, proving that diversity in model selection—specifically pairing models from different labs—is more effective than relying on raw capability. Ghosal concludes that true AI-assisted development requires moving away from passive observation toward active, adversarial verification to ensure code quality and identify genuine failure modes.