Bloomberg’s Automated Uplift generated code-migration requests faster than owning teams absorbed them. Mario Corchero and colleagues report two production campaigns, with 729 direct merges from 2,333 requests.
The system keeps classification, context collection and validation in ordinary code. A model handles a bounded transformation, retries failed checks and describes the result.
- For C++ build migrations, the model extracts structured data. Fixed templates write CMake files, then compilation and unit tests run before review.
- For Python packaging, the model writes changes using source analysis and production logs. Packaging and startup checks leave more runtime risk with the owner.
- C++ requests merged at 34.2% (698 of 2,039), versus 10.5% for Python (31 of 294). Most remaining requests received no developer decision.
- Requests were opened through April 30, 2026, with outcomes measured June 30. They were low-priority maintenance work, without a hard migration deadline.
- Direct merges undercount some assistance: owners later completed migrations in 34 of 51 Python repositories with rejected requests. Fourteen reused most proposed changes.
Caveat: This is an observational report on proprietary systems. Different tasks, populations, support and model versions prevent attributing the merge-rate gap to better checks alone.
The authors call production logs essential, but provide no controlled comparison measuring their effect. Estimated labor savings also come from the team’s judgment.
For maintenance bots, produce evidence that makes a change cheap to accept, and let owners amend the request. Counting generated changes misses the work of landing them.