Can an AI code reviewer reliably review code written by another AI?

[ShipFastFailFast]

One popular workflow is to let one coding agent implement a feature and another agent review it. That sounds efficient, but two models can share blind spots, especially if they rely on the same assumptions.

What checks would make this more trustworthy: independent test cases, a second model, static analysis, security tools, or human review on specific areas?

 

[SaaSnomad]

A different model helps sometimes, but the biggest improvement for me is giving the reviewer the requirements, not the author's explanation.

 

[SideProjectSam]

I'd give the reviewing model the acceptance criteria and ask for adversarial test cases before showing it the patch. Otherwise it tends to politely explain why the first model's design is reasonable. Independent context is important: which users should be allowed to do what, which failures matter, and what the system must never do. After that, I'd still want static analysis and tests that run against actual code, not just a persuasive paragraph from another agent.