Bloomberg's Pomona boosts code quality with tiny AI-powered PRs

Bloomberg researchers have introduced Pomona, a lightweight agentic tool designed for continuous code quality improvement through small, automated changes. Inspired by Kaizen (the Japanese philosophy of continuous improvement), Pomona employs two core agent skills: a Scanning skill that identifies improvement tasks such as linting violations, technical debt markers, and test gaps, and a Repair skill that generates tiny pull requests (PRs) targeting approximately 10 lines of diff. The tool follows a human-in-the-loop design, allowing engineers to review and approve each change, which keeps risks low and maintains trust. This approach contrasts with large-scale refactoring by breaking down improvements into frequent, low-risk increments that can be easily merged.

The team evaluated Pomona through a one-month deployment within a Bloomberg team and a survey of 10 senior engineers. Results were promising: 15 out of 17 generated PRs were successfully merged, with a median time-to-close of under 2 hours—indicating rapid adoption and minimal friction. Additionally, 8 of the 10 surveyed engineers expressed a desire to adopt Pomona in their workflows, specifically praising the small diff sizes and the tool's focused approach to improving code quality. The paper, submitted to arXiv, offers actionable insights for deploying agentic systems in industrial settings, highlighting the effectiveness of small, automated changes for reducing technical debt without overwhelming developers.