Dual-Agent AI System Uses Separate Reviewer to Grade Code Quality Without Metrics

A software developer built a two-agent AI loop to evaluate subjective UI quality, such as whether a region-switching animation feels smooth, which cannot be measured with standard tests. The system uses one Claude Code agent to write and modify code, and a separate read-only agent to grade the output against a fixed, human-written rubric of seven criteria. The reviewer and builder operate in isolated contexts, with the reviewer assessing only screenshots and code rather than the builder's reasoning or self-assessment. In the developer's test run, the reviewing agent rejected the builder's output three times before granting approval, catching a flaw in the test evidence rather than the code itself. The setup deliberately prevents a single agent from both writing and reviewing its own work, a practice the author found leads to near-automatic self-approval.