Single vs. Multi-Agent Systems: Evaluating fault localisation for automated Delphi-to-C# code translation

Sammanfattning

Legacy codebases continue to remain deeply embedded in critical organisational systems such as healthcare, manufacturing, finance, and others. However, maintaining these systems presents an operational threat, for instance, the pool of developers proficient in such systems is shrinking, training costs are rising and compounding technical debt. This study evaluates multi-agent systems for automated code migration with an emphasis on source-to-target code divergence localisation and mitigation. The empirical findings reveal a trade-off: while a multi-agent system improves syntactical correctness, this optimisation comes at the expense of functional correctness, where a single-agent system consistently maintains functional correctness.