AI Code Review Finds More Verified Bugs with a Review Harness

Summary

The article evaluates whether changing the review workflow can improve AI code review without changing the underlying model. A harness runs the model as a team of specialist reviewers rather than relying on a single prompt. Compound Engineering and metareview found 1.6 times as many verified bugs as one-shot prompting and beat the same model in 39 of 42 comparisons. In one result, GLM-5.3 using metareview at low effort achieved a 2.1-times gain. The improvement required about 10 times more tokens, and the harnesses also produced more unsupported findings that still needed checking. The evidence therefore presents orchestration as a potentially important factor in code-review performance, while highlighting verification effort and token use as trade-offs.