World Bank researchers describe a reusable review skill for agent-written research code. It first reads code and documentation without running anything or touching data, then makes only the changes a researcher approves, tracked in Git. In a test on a synthetic package, it flagged all three planted problems as high severity. They also cite a benchmark where letting agents run and revise their code raised correct results from 79% to 97%, with the remaining failures giving wrong numbers and no error. The skill cannot confirm reproducibility or data quality.
