Andrej Karpathy has moved on from vibe coding. The researcher who named the term in February 2025 is now calling it outdated and arguing that the next stage of AI software work is agentic engineering, where the model writes code, runs tests, reads the errors, fixes the code, runs the tests again and reports back when something is shipped.
That shift matters now because Karpathy is not talking about a distant future. In May, he joined Anthropic to rebuild its pretraining research team from the inside, and by February 2026 he was saying the phrase he coined was already behind the curve. For readers searching the term now, the reason is simple: the vocabulary around AI coding has changed faster than the tools themselves, and the person who gave vibe coding its name is the one moving the conversation forward.
Vibe coding was the looser first step. A person typed what they wanted in plain English and let the model build it. It caught on because the first wave of AI coding tools were fast at scaffolding and forgiving on small mistakes, which made it good enough for landing pages in an afternoon and prototypes in a weekend. Agentic engineering takes that one step further. The human gives direction, but the model does more of the work on its own, including testing and fixing what it builds.
That change is already showing up in places that normally move cautiously. In early May, ran a video in which senior engineers at three different companies said the transition had already happened inside their teams. Harvard Gazette covered researchers using agentic engineering to build small internal tools that would have needed a developer to start six months ago. Those were scientists, not coders: they wrote a brief, handed it to an agent and picked up a working tool an hour later.
Even banking, where caution is the rule, has not stayed out of it. PYMNTS reported in May that vibe coding had crossed into regulated banking, and one US neobank shipped a customer-facing feature built entirely by an agent. That is where the friction in Karpathy’s argument shows up most clearly. He is saying vibe coding is already obsolete, yet the same approach was still useful enough to help people ship fast, and in some cases fast enough to turn an idea into a live feature before lunch.
The practical payoff is cycle time. Once teams start running agents, a feature can move from a week to a day, a roadmap can double in scope without doubling the team, and the cost of trying a new idea falls close to nothing. Karpathy’s point is not that software is becoming effortless. It is that the bottleneck is shifting from typing code to directing systems that can already write, test and repair a lot of it themselves. What remains unclear is how far that model spreads outside the companies, labs and banks already using it, and whether the next wave of builders will treat agentic engineering as a novelty or simply as the way software gets made.
