SAFEdit: Multi-Agent Decomposition Resolves Reliability Challenges of Instructed Code Editing

arXiv·medium signalXBlueskyLinkedInCopy link

On EditBench, 39 of 40 evaluated models achieve below 60% task success rate for instructed code editing — highlighting a fundamental gap between code generation and precise editing. SAFEdit introduces a multi-agent decomposition where specialized agents handle intent parsing, localization, and transformation separately, significantly improving reliability. The paper demonstrates that the bottleneck in LLM-assisted code editing isn't model capability but task decomposition — splitting the edit intent from the edit execution yields measurably better results.