Andrej Karpathy: From Vibe Coding to Agentic Engineering

AI development has entered a "Software 3.0" era, where neural networks function as programmable computers and prompting replaces traditional coding. This shift moves beyond mere speed improvements, enabling the creation of entirely new information-processing tasks. While current models exhibit "jagged intelligence"—peaking in verifiable domains like coding while failing at basic common-sense reasoning—they are increasingly capable of autonomous, agentic workflows. Professional software development now requires "agentic engineering," a discipline focused on coordinating these powerful yet stochastic systems without sacrificing quality. As AI automates technical execution, human value shifts toward design, aesthetic judgment, and the fundamental understanding required to direct these agents. Ultimately, while AI can outsource thinking, the human capacity for deep understanding remains the essential constraint in directing complex, agent-driven systems.