Adversarial AI code review: useful on 99 回, but 9× the tokens

Adversarial review works by splitting a diff among several agents with different roles — a conventional reviewer, a bug hunter, and lanes that deliberately hunt for inputs or states that break the change — then cross-checking their claims against the actual code before confirming anything. The team built its version as a Claude Code skill with 8 lanes across three families, plus a separate YAGNI lane that drops or downgrades findings whose real-world harm cannot be shown. Every review logged which lane raised how many findings and how many stuck.

The numbers show both sides of the trade-off. Cross-validation passed 682 findings, and the YAGNI lane cut them to 357 confirmed, so a large share of the raw output never reached a human. The heaviest cost comes before that filter: the main agent must cross-check 1,413 lane findings to reject 1,056 of them, and even discarded findings consume time and tokens. The team also added rules to keep the loop from running away — separating who raised a finding, capping automatic fix cycles, and letting Fable settle split judgments.

Read against the author's conclusion, the practical value seems to depend less on diff size than on what the change touches. A diff under 100 added lines still produced Critical findings in 9 of 23 runs, while nothing was confirmed on any diff under 500 lines in the zero-finding runs — so the deciding factor may be whether the change touches data updates, calculation logic, or authorization. Teams weighing adoption may find the payoff hinges on routing only those critical paths through adversarial review, though the token overhead of cross-validation is likely to remain the main constraint.