Resource leaks cause system performance degradation and crashes. Traditional static analysis approaches rely on predefined rules, limiting their ability to find unknown leaks. Emerging LLM-based methods can locate APIs but face challenges in large-scale projects, including high costs and numerous false positives. To address these limitations, we propose AROP, the first project-level resource leak detection framework integrating LLM-based agents. AROP first extracts project-level metadata to efficiently filter resource-related files. It then employs a detection agent to identify acquisition/release operations, followed by a validation agent and data-flow dependency analysis to reduce false positives. Finally, it determines leaks through leak repair pattern analysis.

We evaluated the effectiveness of our tool on the JLeaks benchmark and 6 large-scale real-world open-source projects. On JLeaks, AROP achieved a recall of 66.1% with a precision of 91.0%, which outperformed the current state-of-the-art, INFERROI, surpassing it by 18.2% in precision while maintaining a comparable recall. In evaluation on real-world projects, AROP discovered 106 previously unknown resource leak defects with a precision of 71.3%. We submitted patches for 28 of these defects, all of which have been confirmed and merged by developers.