Andrej Karpathy, who coined the phrase vibe coding in February 2025, now calls that era over and has adopted a new shorthand: "agentic engineering." The change is more than jargon; it names a distinct workflow in which the AI not only generates code from plain-English prompts but also runs tests, reads failures, fixes its own mistakes and reports back when work is shipped.
Karpathy’s flip matters because vibe coding was simple to describe and fast to use: type what you want in plain English and let the model build it. Founders leaned on it as a near-instant superpower — enough to ship a landing page in an afternoon or a prototype in a weekend. By February 2026, though, Karpathy had publicly moved on from that model and framed the next phase as agentic engineering, where the human role becomes direction rather than hands-on coding.
The mechanics of the two approaches diverge clearly. Vibe coding reduces the friction of turning an idea into code by translating natural language into working snippets. Agentic engineering wraps that translation inside a loop: the model writes the code, runs the tests, inspects errors, revises the code, runs tests again and reports when something is live. That loop transforms AI from a drafting tool into an autonomous operator inside a development pipeline.
Evidence that the shift is already practical, not just theoretical, is beginning to appear. A video published by a major business outlet in early May 2026 showed senior engineers at three different companies saying the move to agentic workflows had effectively happened inside their teams. Reporting in May 2026 further put the method into regulated contexts: one industry outlet found that vibe coding had crossed into regulated banking, and that a US neobank had shipped a customer-facing feature built entirely by an agent.
Academic and research teams are moving in the same direction. Coverage in a university outlet found researchers using agentic engineering to spin up small internal tools that otherwise would have required a developer and half a year of work. Those cases underline the practical consequence: timelines that once measured projects in months can compress to days when the agentic loop operates reliably.
Karpathy’s own movements sharpen the story. After naming vibe coding in February 2025, he joined Anthropic in May 2025 to rebuild pretraining research from the inside. That positional change gives weight to the vocabulary shift; when the engineer who popularized one workflow declares it outdated, product teams and researchers take notice.
The friction in this moment is straightforward. Vibe coding functioned as a bridge and, for many founders, a competitive edge — a quick way to ship without hiring. Declaring that bridge obsolete risks sounding like a loss: if human involvement narrows to giving direction, what happens to the rapid, creative work founders prized? Yet the reporting also treats vibe coding as a transitional phase. It was useful precisely because it let non-developers and small teams move fast while higher-capability systems were still maturing.
That gap — between a valuable interim practice and a broader migration toward autonomous engineering — is the central unresolved question. We have concrete examples of agentic engineering in internal tools, three-company confirmations of internal adoption and a US neobank shipping a feature. But how widely agentic engineering runs in production-grade systems beyond these instances is not yet clear. The next phase will depend on whether models can maintain quality across larger, more complex codebases and whether organizations are willing to give agents responsibility for customer-facing releases.
The near-term consequence is predictable: teams will keep experimenting at the edge, and adopters will demand higher model quality and tighter verification loops. For founders and engineers, the practical choice is not between vibe coding and agentic engineering so much as when to stop prototyping with prompts and start delegating execution and verification to an agentic system. That timing will shape hiring, testing practices and the measure of software risk for the next year.
