Developer mizoko tested three Claude Code review skills over six months, analyzing 1,306 comments. The simplest version achieved a 68.9% fix rate, while the more complex version, which ran 14 perspectives in parallel, reached only 59.2%. This data deserves attention because it offers a concrete counterexample to the assumption that adding more Agents makes AI code review better.

What this is

These are three code-review skills built by mizoko for a mobile app and a backend project. A skill is a reusable instruction package in Claude Code—essentially a dedicated workflow for checking code.

The basic review-pr skill comments line by line on pull requests (PRs), or code change requests. Deep-review invokes as many as 14 sub-Agents in parallel to look for problems from different angles. Nitpick-pr checks only minor details such as formatting. The measurement period ran from April 1 to November 13, 2026.

All three approaches had an incorrect-comment rate of around 10%, but their fix rates differed by nearly 10 percentage points. Deep-review posted an average of 7.8 comments per PR, about 1.8 times the basic version's 4.4. Among those comments, 17.4% contained no mistake but had low priority or fell outside the scope of the change. The result was noise crowding out attention.

Industry view

We believe the most important lesson is not that the simple approach won, but that AI tools cannot be evaluated solely by how many issues they find. We also need to know whether their recommendations are adopted, whether they are accurate, and whether they overwhelm human attention. The basic approach remained because it better matched developers' actual decision-making.

That said, this dataset does not prove that complex approaches have no value. Nitpick-pr achieved a 66.7% fix rate, but it found only formatting issues. Likewise, if deep-review identified a small number of high-severity defects, fix rate alone would not fully capture its performance. There are three obvious limitations: only merged PRs whose authors had responded were counted, which may inflate fix rates; the sources of 36% of the comments could not be confirmed; and the manual determination of whether something had been “fixed” also introduced noise.

We therefore would not read this as an argument for never using multiple Agents. The more reliable conclusion is that teams need deduplication, prioritization and severity-classification mechanisms before adding review perspectives. Otherwise, what they add is usually not insight but another backlog of items to process.

Impact on regular people

For enterprise IT: When enterprises deploy AI code review, they should track fix rate, false-positive rate and false-negative rate, rather than use comment volume to demonstrate that the system works. For teams, saying less while focusing on what matters is generally more useful than applying a comprehensive review everywhere.

For individual careers: Developers' core skills are expanding from “writing code” to “defining review criteria and validating AI suggestions.” People who can judge whether a comment is worth acting on will become scarcer than those who simply maximize generation speed.

For the consumer market: In the short term, code-review agents are more likely to enter enterprise development tools than become mass-market products directly. What users ultimately experience will be faster software updates and fewer defects—or distraction from a large volume of low-value suggestions.