Denys Linkov from Wisedocs presents on benchmarking coding agents when working with new versus legacy codebases, focusing on their experience refactoring a complex AI pipeline in 2025. The company faced challenges including slow throughput, complicated AI pipeline updates, and reluctance to work on a legacy codebase spread across more than ten repositories. To address these issues, they undertook a six-month refactor, consolidating multiple repositories into a monorepo to improve maintainability and scalability, particularly for processing large medical claim PDFs with multiple ML models.
Linkov discusses the concept of technical debt, comparing it to financial debt that compounds over time and emphasizing the importance of evaluating the return on investment (ROI) when taking on such debt. He highlights that while AI engineering has accelerated coding and refactoring speeds, product reliability and code maintainability have not necessarily improved, citing uptime statistics from leading companies. This sets the stage for why their refactor was necessary despite rapid advancements in AI-assisted coding tools.
The team evaluated various open-source orchestrators and built proof-of-concept implementations before committing to the refactor. Early experiments with models like GPT-3 showed promise but required significant manual intervention and produced errors. However, newer models such as Sonnet 4.6 and Opus 4.8 demonstrated substantial improvements, completing tasks faster and with fewer mistakes, enabling more autonomous workflows. This evolution in model capabilities is reshaping software development life cycles, allowing for more complex and longer-running tasks to be reliably automated.
The refactor itself yielded significant productivity gains. By consolidating the codebase into a monorepo, the team accelerated feature development, increased commit rates, and onboarded more developers who were willing to contribute across different areas. Although current large language models (LLMs) like GPT 5.5 can scaffold refactors quickly, they still fall short of fully autonomous, error-free codebase rewrites. Nonetheless, Linkov believes that within six months, models will be capable of handling substantial refactors consistently, as seen in industry examples like Stripe.
In conclusion, Linkov affirms that the refactor was worthwhile despite the rapid improvements in AI coding tools. The monorepo improved code quality, reduced pipeline costs and processing times, and enhanced developer engagement. He advises balancing the timing of refactors with evolving AI capabilities, noting that while future tools will simplify such efforts, legacy codebases can become increasingly difficult to manage without timely intervention. The talk closes with a Q&A addressing practical aspects of the refactor, including repository management, requirement validation, and the evolving role of AI in software engineering workflows.
