Resolving Java Code Repository Issues with iSWE Agent
The new AI agent achieves state-of-the-art resolution rates on Java-specific software engineering benchmarks.
A team of researchers from IBM has published a paper introducing the iSWE Agent, a novel AI system designed to automatically resolve issues in Java code repositories. While most current AI-powered coding assistants and automated issue resolvers, like those based on GPT-4 or Claude, show impressive results with Python, their performance significantly drops for other languages. This creates a major gap for enterprise software development, where Java remains a dominant language. The iSWE Agent specifically targets this underserved area, aiming to bring advanced automation to Java-based projects.
The system's architecture is a key innovation, employing a two-agent framework. One agent is responsible for localizing the bug or issue within the codebase, while the second handles the actual code editing to implement a fix. Crucially, both agents have access to specialized tools built on rule-based Java static analysis and code transformation techniques. This hybrid model—combining the reasoning power of large language models with the precision of traditional, rule-based program analysis—allows iSWE to understand and manipulate complex Java code structures more reliably than model-only approaches.
In evaluations, this methodology proved highly effective. The iSWE Agent achieved state-of-the-art issue resolution rates on the Java-specific portions of two major software engineering benchmarks: Multi-SWE-bench and SWE-PolyBench. This demonstrates a tangible advance in automating a tedious but critical part of the software development lifecycle for a vast ecosystem of enterprise applications. The research suggests that the future of robust AI coding assistants lies not in pure LLMs, but in carefully engineered systems that integrate symbolic and neural techniques.
- Uses a dual-agent architecture with separate agents for bug localization and code editing.
- Integrates novel rule-based Java static analysis tools, creating a hybrid model-based and rule-based system.
- Achieves state-of-the-art performance on Java splits of Multi-SWE-bench and SWE-PolyBench benchmarks.
Why It Matters
Brings advanced AI automation to the massive enterprise Java ecosystem, where current AI coding tools underperform.