Can AI-generated pull requests reduce bugs without increasing hidden technical debt?

Can AI pull requests actually reduce bugs and get merged quickly?

Yes, in controlled settings, AI-generated pull requests can fix bugs efficiently and get merged at high rates. At Bloomberg, a lightweight agentic tool called Pomona generated 39 pull requests over three months, and 32 of them (82.1%) were merged with a median time-to-close of just over two hours [2]. Engineers praised the small diff sizes and focus on code quality, and 10 out of 12 surveyed said they wanted to adopt the tool [2]. This suggests that when AI agents are designed to produce small, focused changes, they can be trusted to fix bugs without overwhelming reviewers.

However, the picture is more nuanced at scale. A large-scale study of 12,433 AI-generated pull requests across 1,495 popular repositories found that rejection odds declined over time, meaning agents got better at producing acceptable PRs [1]. Documentation co-changes were associated with lower rejection odds, and functional failures—especially specification mismatch and logic defects—were the dominant visible rejection pattern among commented rejections [1]. So while AI can reduce bugs, it still struggles with understanding specifications and logic, which are core to bug-free code.

What is the hidden technical debt that AI pull requests create?

The main hidden cost is that AI-generated code often ignores code reuse and introduces redundancy, which silently accumulates technical debt. A study comparing AI-generated pull requests to human-written ones found that LLM agents frequently disregard code reuse opportunities, resulting in higher levels of redundancy [3]. Even more concerning, reviewers tended to express neutral or positive emotions toward AI-generated contributions, masking these quality issues [3]. This means that code that looks plausible on the surface can hide maintainability problems that only surface later.

A deeper analysis of 1,210 merged AI-generated bug-fix pull requests using SonarQube confirmed this pattern: code smells dominated post-merge quality issues, particularly at critical and major severities, while bugs were less frequent but often severe [4]. After normalizing for code churn, the apparent differences in raw issue counts across different AI agents disappeared, showing that larger PRs drive higher issue counts [4]. The key takeaway is that merge success does not reliably reflect post-merge code quality—a PR can be accepted and merged while introducing significant technical debt that will need to be paid later.

How can teams get the bug-fix benefits without accumulating hidden debt?

The evidence points to a design approach that treats AI agents as bounded reasoning tools rather than autonomous committers. The ARC-TD framework proposes a hybrid system that combines static analysis, repository-aware retrieval, dependency graph reasoning, and human governance into a single decision pipeline [5]. This treats code review as a multi-objective optimization problem, jointly estimating defects, maintainability risks, refactoring safety, and debt repayment value [5]. The goal is to transform technical debt management from episodic clean-up into continuous, evidence-grounded remediation.

Practical lessons from the Bloomberg deployment reinforce this: keep pull requests small and focused, use scanning skills to identify and prioritize tasks, and maintain engineer trust through reviewable changes [2]. The study of rejection patterns also suggests that agents should be designed to acquire context, perform local validation, and align with repository expectations—a pattern the authors call a 'reflexive coding agent' [1]. By combining these strategies—small PRs, continuous quality checks, and human oversight—teams can harness AI's bug-fixing speed while avoiding the silent accumulation of technical debt that the larger studies warn about.

About These Sources

This answer is built on 5 studies (4 peer-reviewed, 1 preprint) — published in 2026, 5 from 2024 or later, 1 in Q1–Q2 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 46 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions