exe.dev author argues LLMs are vertically integrated resources, not compilers
The author contends that large language models accelerate decision-making across the stack, marking a shift toward "vibe-engineering" as the future of software development.
An article published on the exe.dev blog challenges the conventional classification of large language models such as Claude as mere code compilers. The author argues that viewing these tools as simple translation layers between natural language and source code is a category error, asserting instead that they function as vertically integrated resources capable of managing strategy, architecture, and coding tasks simultaneously.
The piece details the development of a custom, geographically distributed DNS server designed to resolve latency issues associated with serving all DNS records from Oregon. The project utilised multiple concurrent AI agent loops to handle design, testing, and adversarial code review, resulting in a system that reportedly experienced zero incidents in the month following its launch.
The author notes that the initial question of whether Claude is a compiler was first raised in early 2025, at which point the answer was uncertain. The current conclusion posits that the model operates across the entire stack, from high-level strategic goals down to machine code, without the need for the organisational overhead typically required to bridge these layers in traditional engineering workflows.
Specific technical challenges included managing DNS propagation latency and ensuring consistency across regions. The solution involved building a fully consistent, distributed DNS server where agents researched standard designs, explored failure modes, and planned testing strategies. The final system included unit tests, end-to-end tests, and a shadow-mode for de-risking production rollout.
The project required approximately a week of the author’s attention, during which they read a vanishingly small amount of the actual code. Despite this, the author states they could confidently answer technical questions from colleagues regarding the system’s behaviour under various conditions, suggesting a deep understanding of the architecture despite limited direct code interaction.
The article contrasts this approach with what it terms "vibe-coding," where tasks are handed off entirely to agents. Instead, the author describes "vibe-engineering" as a method that accelerates decision-making and allows engineers to reason about systems at multiple layers. The piece suggests that while some software layers providing only convenience may become obsolete, those enabling the expression of important decisions will remain vital.
The author concludes that while agents are not a free pass to relinquish understanding of deeper system layers, the integration of AI across the engineering stack represents the future of the discipline. The piece ends by asserting that in the near future, this integrated approach to engineering will simply be known as engineering.
