Abstractions for Implementation Are Dead. Long Live Abstractions for Design

The article argues that with AI's ability to generate code, the value of implementation patterns, such as the Strategy pattern in Java, is declining. These patterns were cognitive aids to manage code complexity, but AI can now write such code directly. However, design-level abstractions—decisions about boundaries between domains, data ownership, consistency models, failure isolation, and schema evolution—remain expensive to change and cannot be automated. Examples include choosing between synchronous and asynchronous order processing, defining data owners, and setting consistency boundaries. AI can generate code for any chosen approach, but it cannot make these architectural decisions, which depend on business logic, traffic patterns, SLAs, and operational realities. The article critiques the idea that AI can be asked for design advice, noting that iteration on design is costly and feedback loops take months. It also challenges the idea of specifying all failure scenarios in prompts, emphasizing the engineer's role in anticipating possible scenarios. While a curated knowledge base of postmortems and architecture decisions helps, the final 10% of design judgment remains invaluable, and engineers should focus on these high-value skills rather than on mastering frameworks that AI can handle.