AI agents have bias. The one writing code wants it accepted. The one reviewing wants to find problems. Neither reliably verifies its own conclusions. Adversarial review fixes this: agents find issues, then skeptic agents try to disprove them. Only what survives gets reported.

I ran one on a core library we’ve used for years. What came back wasn’t a long complaint list — it was a small set of real, reproducible issues. Including a few we’d unknowingly built workarounds for.

The review challenged its own conclusions before a human ever had to.

Read my full blog post: Adversarial Review: Making AI Challenge Its Own Conclusions