RefactorPlatform is an open-source evaluation harness for repository-scale refactoring agents. It isolates design choices: model backbone, execution regime (baseline, retrieval-augmented, multi-agent), and prompt specificity. Testing on 100 multi-file tasks showed key results.
AST-aware chunking of code context outperformed naive token-window chunking by 25-30 percentage points. A single retrieval-augmented agent achieved an 86% pass rate, significantly better than a multi-agent configuration (66%). Additionally, retrieval's accuracy gains did not increase the cost per successful refactoring.
This is interesting because it provides a reproducible benchmark for a challenging task and suggests that simpler, retrieval-augmented single agents can be more effective than complex multi-agent systems for refactoring.