vibe coding / agentic engineering / ai

OpenAI-era workflows are splitting into two distinct practices: “vibe coding,” an informal, no-review prompt-and-accept approach useful for rapid prototypes, and a more disciplined professional method recently dubbed “agentic engineering.” The article argues that conflating the two causes real risk—vibe coding breaks down at scale when nobody understands generated code—while agentic engineering describes orchestrating AI agents to implement code under human oversight, with specs, reviews, tests, and ownership of architecture. The author prefers “agentic engineering” (coined by Andrej Karpathy) over “vibe engineering” because it conveys professional legitimacy and clearer responsibilities for teams and hiring. The distinction matters for adoption, risk management, and how orgs structure developer workflows with AI.

AI researcher Andrej Karpathy’s term “vibe coding”—accepting AI-generated code without review—has proliferated and now conflates fleeting prototyping with disciplined, production-safe workflows. The author argues for a clearer term: “agentic engineering,” which Karpathy proposed. Agentic engineering describes orchestrating AI agents that implement, execute, test, and refine code while humans remain architects, reviewers, and decision-makers. That approach preserves engineering discipline—specs, diffs, tests, and oversight—so teams can scale AI productivity gains without the failure modes of unchecked generation. The change in terminology matters because it frames expectations for hiring, governance, and secure, maintainable software development using AI tools.