Vibe Coding Is Already Over. Here Is What Replaced It. And Why It Matters Even If You Are Not a Developer.
Vibe coding changed how software gets built, but agentic engineering is already replacing it. Here is why this shift matters for designers, founders, students, and anyone using AI tools.
In the last two articles, we covered two things.
- —AI is not thinking. It is predicting.
- —AI only knows what you explicitly tell it, not what you mean.
Now we get practical. Because there is a category of AI tools that has changed how software gets built, and understanding it matters whether you write code or not.
"It is called vibe coding. And as of this year, the person who coined the term says it is already obsolete."
Let me explain both halves of that sentence.
What Vibe Coding Actually Was
In February 2025, Andrej Karpathy, a founding member of OpenAI and former Director of AI at Tesla, posted something on social media that named an entire movement.
He described a new way of building software. You describe what you want in plain English. The AI writes the code. You do not read every line. You do not review every change. You accept what it gives you, test if it works, and move on.
His own words were blunt: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."
This was not a small idea. It meant that for the first time, someone with zero coding background could describe an app and watch it get built. Tools like Cursor, Replit, Lovable, and Bolt made this accessible to anyone who could type a sentence.
Karpathy called it vibe coding. The term spread so fast that Collins Dictionary named it their Word of the Year for 2025.
Why It Already Has an Expiry Date
Here is the part most content about vibe coding will not tell you, because most of it was written before this happened.
In early 2026, at a public conversation, Karpathy said something that surprised a lot of people. Vibe coding, in his own words, was already passé.
He described a specific turning point. In November 2025, he was still writing roughly 80% of his own code, using AI for the rest. By December, that ratio had completely flipped. He was delegating 80% of his work to AI agents.
"I can't remember the last time I corrected it," he said. "I just trusted the system more and more."
But here is the important distinction he made. Vibe coding was never meant to be a professional practice. It was designed for what he called "throwaway weekend projects" — quick experiments, prototypes, things where it did not matter if the underlying code was a mess.
The problem is that vibe coding got adopted far beyond that. Startups began shipping production software built this way. And the cracks showed quickly. One widely reported case found that a vibe-coded platform had a security flaw exposing personal data across hundreds of applications. Independent research in 2025 confirmed that while AI got dramatically better at generating code that worked, the security of that code did not improve at the same pace.
Karpathy's response to this was to introduce a new term: Agentic Engineering.
What Agentic Engineering Actually Means
The distinction he draws is precise and worth sitting with.
Vibe coding raises the floor. It means anyone can build something, regardless of skill level. That democratisation is real and valuable.
Agentic engineering raises the ceiling. It is a professional discipline, using the same AI agents, but with oversight, review, and accountability built back into the process.
In his own words: "You are not allowed to introduce vulnerabilities because of vibe coding. You are still responsible for your software, just as before."
The shift is not about the tools changing. The tools are the same — Claude Code, Cursor, and others. The shift is in the relationship between the person and the agent. In vibe coding, you trust blindly. In agentic engineering, you orchestrate deliberately. You stay accountable for the architecture, the security, and the quality, even though you are not personally typing every line.
This is not just a developer conversation. If you are a designer working with AI tools, a founder validating a product idea, or a student building a portfolio piece, the same principle applies. The tool can do the work. You are still responsible for whether the work is good.
The One Thing That Separates a Good Result From a Bad One
Here is a finding that should matter to you regardless of whether you ever write a line of code.
Researchers studying agentic coding tasks in 2025 found something striking. AI agents working without any project context completed tasks correctly roughly 30% of the time. The exact same tasks, with a well-structured context provided upfront, succeeded 90% of the time.
Read that again. Same AI model. Same task. The difference between 30% and 90% success was not the tool. It was the information the tool was given before it started working.
This connects directly to what we covered in Article 2. Every time you open a new session with Claude Code or Cursor, the AI is, in its own engineers' words, "like a new software engineer who has never seen our codebase."
It does not know your tech stack unless you tell it. It does not know your team's conventions unless you tell it. It does not know what has already been tried and rejected unless you tell it.
This is true whether you are building software, structuring a design system, or organising research. The principle is universal. The tool's intelligence is constant. The outcome depends entirely on what you give it to work with.
Why This Matters Specifically For Designers
I would like to address the designers reading this directly.
We are living through the most consequential shift our profession has seen in two decades. I am saying that as someone who has been in design for the last twenty-two years and has seen many shifts.
The AI in Design 2026 report, surveying over nine hundred designers across sixty countries, found something that should stop every design leader in their tracks. Half of the designers surveyed have shipped AI-generated code to production. Not prototypes. Not weekend experiments. Production.
The deliverable of a designer is no longer a Figma file handed to engineering. It is increasingly a working MVP. A functioning prototype. A piece of production code.
"Stay curious. Always."
"Vibe coding became agentic engineering. Agentic engineering will become something else. The field is moving fast."
"The designers who experiment now will move with it. The ones who wait will find themselves trying to learn five things at once, in a profession that has already moved on without them. I have seen this happen before."
This is Article 3 in the series AI From the Inside. Previous: AI Has No Idea What You Mean. It Only Knows What You Said.
Share this article
Related Articles
AI Is Not Thinking. It Is Predicting.
A simple, non-technical explanation of what AI is doing when you talk to it, why it sounds intelligent, and why human judgment still matters.
AI Obedience Is as Dangerous as AI Avoidance
The future of UX will not belong to designers who blindly follow AI, or to those who proudly avoid it. It will belong to designers who can use AI without outsourcing their judgement.