Andrej Karpathy gave vibe coding its name in February 2025 . By February 2026 he had declared it outdated and moved on. In May he joined Anthropic to rebuild its pretraining research team from the inside.

The man who coined the term has already named the next phase. He calls it agentic engineering. If you have been telling clients you vibe code, the phrasing might be a year out of date.

What changed in twelve months

Vibe coding meant typing what you wanted in plain English and letting the model build it. It worked because the first wave of AI coding tools were fast at scaffolding and forgiving on small mistakes. You could ship a landing page in an afternoon and a prototype in a weekend.

Agentic engineering is the next layer up. The model writes the code, runs the tests, reads the errors, fixes the code, runs the tests again, and reports back when something is shipped. Your role becomes direction.

Karpathy's argument is that vibe coding was a bridge phase. The end state was always going to be a system where the human sets the goal and the agent does the work. Bloomberg ran a video in early May where senior engineers at three different companies agreed the change had already happened inside their teams. None of them used the words vibe coding to describe how they ship today.

Harvard Gazette covered the same story from the academic side. Researchers there are using agentic engineering to build small internal tools that would have needed a developer to start six months ago. The people running these experiments are scientists, not coders. They write a brief, hand it to an agent, and pick up the working tool an hour later.

What this means for founders

Founders adopted vibe coding for the same reason they adopted Canva. They wanted to ship faster and without hiring. The tool did the job well enough to feel like a superpower.

Agentic engineering raises the ceiling on what one person can build. You ship working features now. Deployed, with tests, that a paying customer can use. That is a different roadmap. It changes what you can promise. It changes what one person can charge for.

The first thing that changes when you start running agents is cycle time. Building a feature drops from a week to a day, and your roadmap doubles in scope without doubling the team. The cost of trying a new idea drops to almost nothing in the same move. You can spin up an agent to test a feature and pivot before lunch. Knowing how to break a project into clean tasks and review the output is worth more than knowing how to write the code yourself.

PYMNTS reported in May that vibe coding had already crossed into regulated banking, with one US neobank shipping a customer-facing feature built entirely by an agent. The compliance team approved it after a four-hour review. A year ago that build would have taken four months. The regulators are catching up to a model that already exists in production.

Beyond vibe coding: how to get AI agents to do the work

Pick one project on your roadmap that has been stuck. The one you keep meaning to ship and keep deprioritising because it would take a week you do not have. Open Claude Code, Cursor, or another agent-based tool.

Write the spec as if you were briefing a freelancer. One paragraph on what it should do. Three bullet points on what good looks like. One sentence on what it should not do. Run the agent. Read what it produces. Send it back with one correction. Run it again.

You will hit the limits within an hour. That is the point. You learn agentic engineering by directing one and watching where it gets stuck. The patterns you see in those failures are what your next prompt fixes. The skill compounds quickly once you stop treating the agent like a search bar.

Once you've learned from your first round, set the target on what you'll build next. Track which categories of task the agent finishes on the first attempt, which need one round of corrections, and which need you to take over. That spreadsheet is your map of where agentic engineering pays off in your business.

The founders who run this loop fifty times this quarter will know things in October that other founders will not know until next year.

Agentic engineering is the future

Karpathy moved to Anthropic for a reason. The model is the bottleneck on what an agent can do. The companies building the best models are pulling ahead because their tool finishes more of a task without a human stepping in to correct. Pick your agent tool based on the model under it. The agent that finishes the most work on the first attempt is the one worth paying for. Change tools when the model under it stops winning.

The vibe coding era taught you that you can build things you do not understand. The agentic engineering era will teach you something harder. You can ship things you do not understand. That is the real bet you are making this year. Be ready for it.