SDeCopilot: Assisting Developers in Building Correct Software with Debugging Agents

• Applicant: Dr. Cedric Richter
• Subject Area: Software Engineering and Programming Languages ; Methods in Artificial Intelligence and Machine Learning
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 571455992

Software developers are increasingly relying on Artificial Intelligence (AI) based coding assistants to perform everyday coding tasks. Although this promises a significant boost in productivity, recent studies suggest that the use of AI coding assistants could also lead to a significantly higher number of software bugs in code. Software developers are therefore often plagued with fixing AI-generated bugs, making the task of creating correct and reliable software even more difficult than before. This observation then raises the question whether we can actually use AI to support software developers in building correct and reliable software systems. We see potential for addressing this question in the intersection of AI and traditional software engineering (SE) methods. Recent evidence suggests that AI systems can generate verifiable claims about code, such as potential candidates for test specifications, loop invariants, or functional contracts. Although these claims may not be correct, they can still serve as a valuable starting point for uncovering inconsistencies or unexpected behaviors. Traditional SE methods such as fuzzing, testing, formal verification, or static analysis are more effective in providing reliable insights into the program behavior. However, to find incorrect behavior, these methods often require some form of correctness specification, which is typically provided by a software developer. This reliance on human-provided specifications creates a bottleneck that AI could likely alleviate by suggesting plausible specification candidates. These candidates can then be used by traditional SE methods to identify potential inconsistencies or software bugs. In this Walter Benjamin project, we aim to systematically research how the capabilities of modern AI systems can be effectively combined with traditional SE methods for automatic bug discovery and program repair. Ultimately, we envision a novel software debugging copilot (SDeCopilot), an AI debugging system, which utilizes existing SE methods to autonomously (1) scan given code repositories, detect and report potential issues (e.g., inconsistencies with existing code or documentation), and (2) suggest fixes for the detected issues.

• DFG Programme: WBP Position