Agentic engineering is what vibe coding turned into once the novelty wore off. Andrej Karpathy, who coined vibe coding in February 2025, renamed his own practice within a year: the defining trait isn’t typing less code, it’s that the human stops writing the implementation directly and instead defines the goal, breaks it into a spec or a set of roles for different agents, and reviews the result against an explicit standard rather than accepting whatever comes back.
The shift matters because the failure modes are different. Vibe coding fails silently, a wrong assumption gets accepted because nobody read the diff closely enough. Agentic engineering fails loudly, or should, because a verification step is built into the process before an agent’s output ships, the same discipline a test suite provides for human-written code, applied one layer up.
The tell that separates the two in practice: if review only happens after the code is generated, that’s still vibe coding with extra steps. If the acceptance criteria and the decomposition into roles happen before an agent writes a line, that’s agentic engineering.