As AI coding agents become ubiquitous in development workflows, data scientists face a critical challenge: knowing what to look for in auto-generated code. This article argues that while AI can generate code rapidly, human expertise in reviewing and guiding these tools has become more valuable than ever. The shift from writing code to reviewing it represents a fundamental change in how developers work.

Drawing from established software engineering practices, the piece focuses on three core concepts: code smells, abstraction, and design patterns. These aren't new ideas, but they've gained renewed importance as coding assistants handle the manual labor of writing. Seasoned developers instinctively recognize problematic code patterns through years of experience, but junior practitioners now need to develop these skills to effectively collaborate with AI tools. The article positions this knowledge as essential for anyone wanting to excel in today's AI-augmented development environment.

The piece uses a practical example from machine learning engineering to illustrate divergent change - when a single class handles too many responsibilities. A `ModelPipeline` class that combines data loading, cleaning, and training creates three potential failure points. By coupling platform, data engineering, and ML concerns, developers multiply the risk of bugs and context pollution for AI agents. The solution involves separating these concerns into distinct classes, reducing operational risk and making the codebase more maintainable.