AgentFL is a novel agent-based framework that enables large language models to identify bugs across entire projects, overcoming the context limitations of traditional LLM-based fault localization.

• Uses a divide-and-conquer approach with multiple specialized agents to handle project-level context
• Achieves 3.67× improvement over state-of-the-art methods in project-level fault localization
• Demonstrates 96.3% effectiveness in accurately identifying buggy files
• Employs an innovative collaboration mechanism where agents share knowledge to gradually narrow down the location of bugs

This research significantly advances automated debugging capabilities for engineering teams working with large codebases, potentially reducing debugging time and increasing developer productivity in real-world software development environments.

AgentFL: Scaling LLM-based Fault Localization to Project-Level Context