Refactoring Runaway: Coding Agents Inherit Tangled Refactoring from Training Data

arXiv·high signal

Analysis of 3,691 valid patches from Multi-SWE-bench reveals that LLM-based coding agents (including SWE-Agent variants) systematically produce tangled refactoring — mixing bug fixes or feature changes with unnecessary refactoring inherited from training on open-source repos. The paper identifies this as a fundamental behavioral pattern, not a prompting failure, and proposes mitigation strategies. Practitioners using AI coding agents should expect 20-40% of generated patches to contain scope creep that complicates review and introduces regression risk.