AGENTFL is a multi-agent system based on ChatGPT for automated fault localization at the project level. It addresses the limitations of existing LLM-based approaches in handling large codebases by decomposing the localization process into three stages: Fault Comprehension, Codebase Navigation, and Fault Confirmation. Each stage is handled by specialized agents with different expertise, using strategies like Test Behavior Tracking, Document-Guided Search, and Multi-Round Dialogue to overcome challenges in each step. AGENTFL outperforms existing LLM-based approaches in localizing bugs, achieving 157 out of 395 bugs within Top-1 on the Defects4J-V1.2.0 benchmark. It also shows complementarity with learning-based techniques. An ablation study confirms the importance of its components, and a user study demonstrates its practical usability. Cost analysis shows AGENTFL is efficient, requiring only 0.074 dollars and 97 seconds per bug. The system is designed to mimic human debugging practices, enabling more controlled and effective fault localization. AGENTFL's approach enhances the capabilities of LLMs by breaking down the localization process into manageable steps and leveraging multiple agents with specialized tools. The system's effectiveness is validated through extensive experiments on real-world Java projects, showing its potential for practical use in software debugging.AGENTFL is a multi-agent system based on ChatGPT for automated fault localization at the project level. It addresses the limitations of existing LLM-based approaches in handling large codebases by decomposing the localization process into three stages: Fault Comprehension, Codebase Navigation, and Fault Confirmation. Each stage is handled by specialized agents with different expertise, using strategies like Test Behavior Tracking, Document-Guided Search, and Multi-Round Dialogue to overcome challenges in each step. AGENTFL outperforms existing LLM-based approaches in localizing bugs, achieving 157 out of 395 bugs within Top-1 on the Defects4J-V1.2.0 benchmark. It also shows complementarity with learning-based techniques. An ablation study confirms the importance of its components, and a user study demonstrates its practical usability. Cost analysis shows AGENTFL is efficient, requiring only 0.074 dollars and 97 seconds per bug. The system is designed to mimic human debugging practices, enabling more controlled and effective fault localization. AGENTFL's approach enhances the capabilities of LLMs by breaking down the localization process into manageable steps and leveraging multiple agents with specialized tools. The system's effectiveness is validated through extensive experiments on real-world Java projects, showing its potential for practical use in software debugging.