Vercel Tech Lead: AI is Distilling Software Engineering, Not Replacing It

Javi Velasco, a Tech Lead at Vercel, presented a compelling perspective on the evolving role of software engineers in the age of Artificial Intelligence, arguing that while AI automates mechanical coding, it amplifies the human demand for critical context and design. Citing recent industry observations, Velasco noted a growing trend where developers, including himself and even the CEO of Spotify, report writing significantly less raw code. This shift, however, comes with its own challenges, as evidenced by Amazon’s recent increase in P1 incidents, many attributed to AI-generated code. Velasco posits that AI, despite its generative capabilities, fundamentally lacks the deep system understanding, domain knowledge, and architectural context crucial for producing robust, production-ready software, often leading to performance issues or unintended side effects.

This contextual gap underscores the “Human in the Loop” principle, where engineers become indispensable in providing the intricate details—business rules, technical constraints, non-functional requirements, and past design decisions—that AI cannot infer. Velasco illustrated this with examples from his own projects, where AI struggled with product relevance, scalability (e.g., handling 10 million items vs. 30), and complex refactoring, even when explicitly warned about side effects. He stressed that AI isn’t curious and often “doesn’t know what it doesn’t know,” perpetuating errors with unwarranted confidence. The quality of an AI’s output, therefore, is directly proportional to the quality of the context it receives, making practices like clear abstractions, well-defined types, and clean architecture—once considered aesthetic—now critical investments against “context poisoning.”

Velasco advocates for a new development paradigm that integrates AI as a powerful tool for code generation, but firmly places humans at the helm for design, review, and strategic guidance. He outlined a workflow involving an “AI Software Architect” agent for high-level planning and an iterative local review process, facilitated by a custom diff tool, to meticulously scrutinize and refine AI-generated code before deployment. This approach, he contends, distills the essence of software engineering into its most valuable and challenging components: understanding problems, modeling domains, and designing resilient solutions. For aspiring engineers, Velasco advises prioritizing classic software design principles over mere code generation, emphasizing that the ability to provide deep context and critically evaluate AI’s output is the true currency of programming in 2026 and beyond.