Pomona: Continuous Code Quality Improvement via Small, Automated Changes at Bloomberg
This paper introduces Pomona, a lightweight agentic tool inspired by Kaizen that continuously improves code quality at Bloomberg by automating the discovery and incremental repair of technical debt through small, low-risk pull requests, achieving high merge success rates and strong engineer adoption in a one-month deployment.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are managing a massive, bustling library (your company's codebase). Over time, books get slightly misshelved, some pages have coffee stains, and a few notes are written in pencil that need to be erased. Usually, the librarians (the software engineers) are too busy writing new books to fix these small, nagging issues. They know the library needs tidying, but they don't have the time to do it all at once.
Enter Pomona, a new, helpful robot assistant designed specifically to do the "little things" that humans keep putting off.
Here is how Pomona works, broken down into simple concepts:
1. The Two-Step Dance: Scanning and Repairing
Pomona isn't a magic wand that fixes everything at once. Instead, it works like a diligent janitor with a specific routine, using two main "skills":
The Scanner (The Eyes): This part of Pomona walks through the library looking for small problems. It checks for things like:
- Books left on the wrong shelf (linting errors).
- Sticky notes saying "Fix this later" that have been there for years (TODOs).
- Pages that are blank and useless (dead code).
- Missing chapters in the index (test gaps).
Once it finds these issues, it writes them down in a Backlog (a to-do list), sorting them by how important they are and how easy they are to fix.
The Repairer (The Hands): This part picks the very first, easiest, and most important item from the to-do list. It then makes a tiny, tiny fix—think of it as rearranging just 10 lines of text or fixing a single typo. It doesn't try to rebuild the whole library; it just fixes one small shelf.
2. The "Tiny Pull Request" (The Human Check)
Here is the most important part: Pomona doesn't just change the books on its own. It creates a Pull Request (PR).
Think of a Pull Request as a sticky note Pomona leaves on a book saying: "Hey, I noticed this page was upside down. I flipped it. Please take a quick look and tell me if I did it right."
- Small is Safe: Because the changes are so small (about 10 lines), a human librarian can glance at them in seconds.
- Human in the Loop: The human must say "Yes" before the change is made. This keeps the engineers in control and builds trust. They aren't afraid the robot will accidentally delete the whole library.
3. The Kaizen Philosophy
The paper compares Pomona to Kaizen, a Japanese philosophy of "continuous improvement." Instead of trying to do a massive, scary renovation project once a year, Pomona suggests making tiny, constant improvements every day. It's the difference between trying to clean the whole house in one weekend versus spending 10 minutes a day tidying up one room.
4. What Happened When They Tried It?
The team at Bloomberg (a huge financial company) let Pomona work for one month. Here is what they found:
- It Worked: Out of 17 tiny fixes Pomona suggested, 15 were accepted and merged into the code.
- It Was Fast: Most of these fixes were approved and closed in less than 2 hours.
- People Liked It: When the engineers were asked if they wanted to use Pomona, 8 out of 10 said yes. They loved that the changes were small and easy to read.
- The Bottleneck: The only time things got stuck was when Pomona got a little too excited and tried to fix the same thing twice before a human could check the first time. The team fixed this by telling Pomona to check if a fix was already in progress.
5. Why This Matters
The paper concludes that while AI is great at writing big, complex code, it is currently best at being a helpful assistant for small, boring tasks.
- Trust: Engineers are skeptical of AI making big changes, but they trust it with small, obvious fixes.
- Focus: By letting the robot handle the "trash picking" (dead code, old notes), humans can focus on building new features.
- Control: The system works because the human always has the final say.
In short, Pomona is a tool that helps teams keep their code clean and healthy by making small, automated, and easily approved improvements, rather than waiting for a human to find the time to do it all manually.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.