An AI review agent analyzes proposed changes and flags potential issues for human judgment. It can find problems that other checks miss, but a plausible finding is not proof of a bug: the agent may lack the requirements or architectural context needed to distinguish a defect from intended behavior. section 8.2.3 section 10.5.1
Reviewing a markup parser
An AI reviewer flagged newline matching, replacement order, nested parentheses in links, and missing HTML escaping. The escaping issue was also flagged by static analysis; the newline issue was not. But whether bold-italic replacement is wrong depends on the markup requirements, which the reviewer did not know. section 8.2.3
A reviewer may also be less independent than it appears. Research on LLM evaluation of general text found that models favored their own outputs, with follow-up work linking preference to how predictable the text was to the judge. The research did not measure code review specifically, so it is a reason for caution rather than a demonstrated code-review error rate. section 8.2.3
Use AI review as one layer in a review pipeline:
- Encode known, enforceable rules in lint configuration. Point coding agents to those rules so they can check their changes before submission. section 10.5.1
- Run the review agent as a first pass on pull requests. Use its findings to triage the human queue, not to replace human review of intent, trade-offs, and strategic alignment. section 10.5.1
- Use policy-as-code in CI to enforce compliance rules, and check AI findings against the actual requirements and team conventions. section 10.5.1 section 8.2.3
An integrated review agent is also part of the cost of an AI-assisted workflow. Evaluate its contribution alongside the subscriptions, usage charges, and less visible costs of that broader workflow. section 9.1

