Exploring aspects of self-evolution of software systems with agents in industrial contexts.


Motivation 
Automatic refactoring will become increasingly important with AI-supported software engineering. As AI is also used for large projects, we must investigate how to enable AI systems to handle large codebases and, therefore, large-scale refactorings. By large-scale refactorings, we mean refactorings that need changes to multiple source code files. While the research would start with mid-sized projects, we would also like to investigate refactorings for 300KLOC+ systems. Furthermore, refactorings as well as requirements-based application generation have to move in the direction of brownfield industry projects. This project will investigate several aspects of agent usage in this context.

Aim 
Development and evaluation of an agent system that is capable to perform (certain) refactorings with minimal (ideally no) support or input from the software engineer. The minimality of interaction with the software engineer is not just measured, but is a requirement equally important as the correct refactoring, and is therefore explicitly considered in our approach. Furthermore, we will concentrate on large-scale refactorings from various viewpoints: Exploring techniques for symbol and dependency management in large systems, exploring successful use of small LLMs as well as self-learning approaches.

 

Duration 10/2026- 09/2027

Research Partner  Siemens AG

Contact Reinhold Plösch

SEACT – Self-Evolution with Agents – Concepts, Techniques and Experimental Data