SGAgent: Suggestion-Guided LLM-Based Multi-Agent Framework for Repository-Level Software Repair

A research team from Nanjing University and the Chinese Academy of Sciences has introduced SGAgent, a novel multi-agent framework that significantly advances automated software repair. The system addresses a critical gap in existing AI repair tools by implementing a three-phase 'localize-suggest-fix' paradigm, where a specialized 'suggester' agent bridges the reasoning gap between identifying bugs and generating fixes. This architectural innovation, combined with a repository knowledge graph for enhanced contextual awareness, enables SGAgent to achieve state-of-the-art performance on challenging benchmarks like SWE-Bench, where it demonstrates 51.3% repair accuracy using Claude 3.5 Sonnet as its base model.

SGAgent's technical architecture features three specialized sub-agents: a localizer that identifies buggy files and functions with 81.2% and 52.4% accuracy respectively, a suggester that incrementally retrieves relevant context to understand bugs, and a fixer that generates final patches. The framework's cost efficiency is particularly notable at $1.48 per instance, making it economically viable for real-world deployment. Beyond standard bug repair, SGAgent shows strong generalization with 48% accuracy on vulnerability repair datasets VUL4J and VJBench, indicating potential applications across security domains and multiple programming languages. This represents a significant step toward practical AI-assisted software maintenance at the repository level.