The author reflects on the significant impacts of AI coding tools, particularly large language models (LLMs), on software engineering as 2026 approaches. The cost of producing high-quality code has dropped considerably, shifting the bottlenecks in software engineering to areas beyond coding itself. While building and evolving software systems have become more efficient, operating them still faces challenges that LLMs have not yet fully addressed. Companies will need to adapt, especially within product-oriented teams that stand to gain more from these tools compared to infrastructure teams. Key shifts expected in 2026 include a focus on infrastructure abstractions, where rapid deployment and rollback of binaries will be essential. The quality and speed of continuous integration (CI) infrastructure will become even more critical as AI begins to write more code. There's also a need to rethink testing strategies, moving beyond just unit tests to include property testing and formal verification. Human-guided abstractions will be vital to prevent LLMs from generating low-quality code, emphasizing the importance of clear module boundaries and interfaces to maintain code quality. Human code review will evolve into a more critical bottleneck, requiring engineers to develop a new sense of “review taste” as they increasingly rely on LLMs for code generation. This creates a paradox for junior engineers who need experience in writing code to develop this intuition but may find themselves writing less. The author anticipates increased variance in project timelines as tasks become more influenced by how well they can be automated with LLMs. While the falling costs of code may shift some "build vs. buy" decisions, the impact on infrastructure and compliance services will remain limited compared to simpler software solutions. Several open questions remain, including the necessity of human review for every line of code and the implications of significantly faster and cheaper AI models.