I was halfway through writing a post called "How to Vibe Code Faster" when Andrej Karpathy retired the term.
The draft was open in one window. His tweet, replacing vibe coding with "agentic engineering," was open in another. I read it twice, then closed the draft without saving. For about ten minutes I sat there wondering if my entire content calendar had just been quietly disassembled by a single guy in California with a very good track record for naming things.
Then I reread the tweet a third time and realized he had done something more interesting than killing a word. By promoting agentic engineering, he had accidentally explained why vibe coding is not actually dying. It is splitting in half.
Here is my attempt to articulate why that matters, and why I think Karpathy is exactly half right.
The tweet that retired the word
On the first anniversary of coining "vibe coding," Karpathy posted what I can only describe as a eulogy for his own meme. In Karpathy's anniversary tweet, he noted that the term had become a catchall for everything anyone did with an AI coding assistant, from a ten-year-old shipping a Roblox mod to a staff engineer at Stripe orchestrating a fleet of agents across a monorepo. The word had collapsed under its own weight.
His replacement was "agentic engineering." The framing was precise. Where vibe coding described a posture, accepting AI output on feel, agentic engineering described a practice: planning, delegating, verifying, and composing the work of autonomous agents into something coherent.
The tech press picked it up within hours. The New Stack ran a piece titled "Vibe Coding Is Passe", quoting developers who said they had been waiting for someone with Karpathy's authority to name what they were already doing. Bloomberg called it the Great Productivity Panic. I watched a friend last week run six Claude Code sessions at once, each on its own branch, and he looked less like a coder and more like a conductor who had lost the score.
So the word is out. Long live the word.
Except it isn't, and the reason is the part nobody is talking about.
What agentic engineering actually means
Strip the term of its novelty and look at what it describes. Agentic engineering is the practice of treating AI agents as junior collaborators you manage rather than tools you operate. The human role shifts from typing to something closer to running a small, very fast, very literal engineering team. This is closer to running a kitchen during dinner service than writing code. You are not at the stove. You are calling tickets, tasting sauces, watching which station is about to fall behind.
You plan the work in advance. You write specifications the agents can read. You break tasks into parallel workstreams. You review the diffs, not the keystrokes. You build verification harnesses that can catch agent failures without your eyes on every line. You learn which agents are good at which tasks the same way you learn which teammates are good at which tasks, and you route work accordingly.
This is a real skill. It takes months to develop. It requires familiarity with how different models fail, how to write prompts that survive long context windows, how to design test suites that can run unattended, and how to sequence work so that one agent's output feeds cleanly into another agent's input. None of this is what a person means when they say, "I vibe coded a landing page over coffee."
Agentic engineering is, functionally, a new discipline. It sits closer to systems engineering than to writing code. The people doing it well are mostly senior engineers who already knew how to plan work, who have just gained a force multiplier. I wrote earlier that vibe coding is the new product management, and agentic engineering is the extreme version of that thesis. You are no longer implementing. You are directing implementation at scale.
It deserves its own name. Karpathy is right about that.
Why the rebrand is half right
The most charitable reading of Karpathy's move is that he is protecting the field from its own success.
Here is the problem. When a term like vibe coding catches fire, it attracts a long tail of usage that the original meaning cannot carry. A venture capitalist starts saying their portfolio companies are "all vibe coded." A CTO claims her team "vibe codes in production." A CEO promises investors that "vibe coding" will cut engineering headcount in half. At some point the word stops describing what anyone is actually doing and starts describing what people wish were true, which is the point at which it becomes useless for people who need to think carefully.
Naming the mature practice "agentic engineering" is a way of pulling the serious work back to shore. It tells recruiters that there is a hireable skill here. It tells skeptics that the practice has grown up. It tells practitioners that they belong to a discipline, not a trend. A year of field experience has taught the people at the top of the practice that what they are doing is not casual, and they want a word that reflects the care they take.
The rebrand also serves a more honest purpose. It acknowledges that the original term, with its implied shrug, was never going to sit well with engineers who take craft seriously. You cannot build reliable software by vibes alone. The people orchestrating multi-agent workflows on real codebases know this in their bones. They needed a word that sounded like work, because the work is real.
On all of that, Karpathy is correct. The top of the practice has matured, and the top of the practice deserves its own vocabulary.
But the top of the practice is not the whole practice.
Why the rebrand is half wrong
Here is the part the rebrand leaves behind.
For every engineer running a Claude Code swarm across a microservices monorepo, there are a hundred people who just opened Cursor for the first time last week and typed, "make me a habit tracker," and were delighted when something usable appeared. For every principal engineer writing agent specs in YAML, there is a founder in a coffee shop building her first MVP by describing it in plain English to an AI that keeps guessing what she means and keeps being roughly right.
These people are not doing agentic engineering. They do not have a plan. They do not have a test harness. They do not have specifications. They have an idea and a tool and a feeling of momentum. They are vibe coding in the original, literal sense of the word, and the term still fits them like a glove.
Trying to rename what they do "agentic engineering" is like telling a weekend home cook they are actually doing "culinary systems design." The words are technically accurate. They also miss the point. The home cook wanted a word that gave her permission to be imprecise, to trust her senses, to move fast and fix things later. That is what vibe coding gave to millions of people who would never have touched a text editor otherwise.
The term did not just describe a practice. It gave a whole cohort of non-engineers permission to join software. I wrote a primer on what vibe coding is last year, and the single most common response I got was from people who said the word itself had unlocked something for them. They had wanted to build things for years. They had been intimidated by the rituals of traditional software engineering. And then a word arrived that said: it is fine to not know what you are doing. You are allowed to feel your way through.
You cannot retire that word without taking something away from those people. And the replacement, "agentic engineering," is the wrong gift. It sounds hard. It sounds professional. It sounds like a job posting at a Series B startup. It closes a door that vibe coding opened.
Run the napkin math. If vibe coding unlocked software for, say, ten million non-engineers in a year, and agentic engineering is being practiced seriously by maybe fifty thousand power users, the ratio is two hundred to one. You do not rename the thing two hundred million hours of practice are pointing at to please the fifty thousand people at the top.
The two-tier future
Here is what I think is actually happening, and why I do not think Karpathy's rebrand will stick the way he wants it to.
The practice is diverging into two tiers, and both are legitimate.
At the top, you have agentic engineering. Senior practitioners running multi-agent workflows, maintaining production systems, writing specs and verification harnesses, billing the kind of rates that reflect the leverage they are pulling. This is the world Karpathy is describing. It is real, it is growing, and it needs a name that distinguishes it from the hobbyist tier. The people at this level read guides like 50 Claude Code tips the way chefs read technique manuals: closely, skeptically, looking for the one detail that upgrades their whole operation.
At the bottom, or maybe at the other end (I do not love the vertical framing, but I am stuck with it for now), you have vibe coding in its original form. A much larger population of people using AI to build things they could not have built before. Small utilities. Personal sites. Weekend projects. First products. Internal tools at non-technical companies. The vibe coder is not pretending to be an engineer, and that is the whole point. They are using software the way previous generations used spreadsheets: as a thing that extends what they can do on their own, without needing a specialist.
Both of these practices use the same underlying tools. The Cursor window looks identical whether a staff engineer or a teenager is sitting in front of it. But the mental model, the expectations, the standards of success, and the required skill are all completely different. Treating them as one thing was always going to fail. That is what Karpathy's rebrand is correctly responding to.
Treating them as if only the top tier deserves a name is the mistake.
If the rebrand sticks at the top and the original term survives at the bottom, the field will actually be healthier. Agentic engineering will mean something specific and hard. Vibe coding will mean something approachable and forgiving. Recruiters will know what to look for. Beginners will know where to start. The press will stop using one phrase to describe both a twelve-year-old's Minecraft mod and a production deployment at a fintech.
That future is the one I am rooting for. One word cannot carry the whole practice. Two words, each carrying half, might.
Karpathy did not actually say vibe coding was dead. He said it was a first draft. He has more credibility on this than anyone, so when he puts a new word on the table, people listen. But the word that wins is not always the word the expert picks. The word that wins is the word the largest group of users needs. I do not know yet whether agentic engineering will hold. It might also become one of those phrases that shows up on LinkedIn bios for a year and then quietly disappears while the underlying practice keeps growing under whatever name works. Vibe coding, meanwhile, is not going anywhere, because the people it describes are not going anywhere, and they are not suddenly going to start calling themselves agentic engineers just because a famous researcher said so.
Maybe the right move, for those of us who write about this stuff, is to hold both words at once for a while and see which ones the people doing the work actually use when nobody is listening. If you are running a multi-agent workflow or building your first app from a coffee shop, I would love to know which word feels like yours, and whether either of them does.
