PI: Christoph Reichenbach (CR), Dept. of Computer Science, Lund University
co-PI: Marco Kuhlmann (MK), Dept. of Computer Science, Linköping University
A single large-scale change to a software system may affect hundreds of thousands of lines of source code, yet a senior software engineer can often compress the essence of the change into a few sentences of natural language. Coding agents can mimic this ability and invert it: using large language models (LLMs), they expand a developer request into a series of edits, combining background knowledge from their training data with the source code at hand to contextualise the requested change. However, they lack the correctness guarantees of classic software tools (such as compilers), which impairs their reliability in practice. We propose to address this limitation by fitting coding agents with a formal safety net that we originally developed for program metamorphosis, a technique for manual incremental software evolution with strong behavioural guarantees. This safety net is designed for providing feedback that enables and steers incremental program updates, and we expect it to not only prevent accidental behavioural change but also to help guide the agentic transformation process.
Project number: H7