Bloomberg preprint reports 729 successful migrations using AI-assisted pipelines

Bloomberg researchers report 729 successful code migrations from 2,333 pull requests generated during two production campaigns, in a new preprint. The work covered approximately 1,900 candidate repositories for Python packaging changes and approximately 6,700 for a C++ build system migration. The pipelines used deterministic processing wherever possible and confined language models to bounded file transformations, informed by execution logs and usage traces. Reviewing, testing and deploying changes proved a greater bottleneck than generating them. Verification before opening pull requests, including C++ compilation and testing, reduced those costs; uniform repositories also allowed fixes to carry across projects.