AI review loops don’t always stabilise
Naively alternating AI code review and AI-applied fixes can increase defects instead of converging: model opinions vary between runs, reviews expand scope, and false positives or hallucinations introduce new bugs.
The author’s small experiment asked an agent to write a non-trivial CIDR radix-trie library, then alternated delegated reviews and fixes for three rounds. Defect count increased each round. The linked write-up shows why the missing acceptance criteria—what “perfect” means, and which tests define it—dominated the outcome; later reviewers also received less change history to test whether they audited the artifact rather than the narrative. The Lobsters page exposed no usable comment text during review.