Vibe Coding Explained: What It Is, Where It Came From, and What It Means for Developers

One Tweet That Changed How Developers Talk About Their Work

In February 2025, Andrej Karpathy — co-founder of OpenAI, former AI Director at Tesla, one of the most respected researchers in the field — posted a tweet that went viral:

*"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs are getting too good."*

Within weeks, "vibe coding" was everywhere. A year later, it has its own Wikipedia page, its own controversies, and its own successor term. Here is what it actually means — and what it tells us about where software development is heading.

What Vibe Coding Actually Is

Vibe coding is a development approach where you describe what you want to build in plain language, let an AI model (Cursor, Claude, GPT-4) generate the code, and iterate by describing changes rather than writing them.

You are not reviewing every line of code. You are not necessarily understanding the implementation details. You are directing — describing outcomes, testing results, asking for changes — and trusting the AI to handle the actual writing.

The "vibe" in vibe coding refers to this feeling: you are coding by intuition and direction rather than by deliberate construction. You catch the vibe of what you want, communicate it, and the AI builds it.

What It Looks Like in Practice

Instead of:

Writing a React component from scratch, looking up the API syntax,
handling the state management, debugging the CSS layout

Vibe coding looks like:

"Build me a card component that shows a blog post title, date, and
excerpt. Make it responsive. Add a hover effect."

→ AI generates the component

"The hover effect is too slow. Make it 150ms."

→ AI updates one line

"Add a read time badge in the top right corner."

→ AI adds the badge

The difference is not just speed. It is a fundamentally different relationship with the code — you are the director, the AI is the implementation layer.

Who Started It and Why It Spread

Karpathy is not a typical tech influencer. He is the person who built Tesla's Autopilot vision system and was one of OpenAI's founding researchers. When he says AI coding tools have gotten good enough to change the fundamental nature of software development, the industry listens.

The timing also mattered. By early 2025, tools like Cursor (AI-native code editor), GitHub Copilot, and Windsurf had reached a level where multi-file code generation, context-aware completions, and conversational iteration were genuinely reliable. The capability had arrived at the moment Karpathy named it.

According to 2026 statistics, 92% of US developers now use AI coding tools daily and 82% of global developers use them at least weekly. Vibe coding, or something close to it, is not a fringe experiment — it is the mainstream.

The Honest Criticisms

Vibe coding has real problems, and the developer community has been vocal about them.

You May Not Understand What You Shipped

If you accepted AI code without reading it, you do not understand what you deployed. You cannot debug it when something breaks at 2am. You cannot explain it to a colleague. You cannot extend it confidently. You are the owner of a system you did not build and cannot fully account for.

Security Vulnerabilities Hide in Unread Code

AI models generate code with security vulnerabilities. SQL injection, improper authentication, exposed API keys in client-side code — these appear in AI-generated output regularly. A developer who reads the code catches them. A developer who accepted the vibe misses them.

Technical Debt Accumulates Invisibly

AI-generated code tends toward the immediately functional rather than the well-structured. It solves the stated problem without considering the system's long-term architecture. Enough vibe coding in a codebase produces something that works today and becomes increasingly difficult to maintain over time.

The Gap Between Demo and Production

Vibe coding produces impressive demos quickly. Production systems have error handling, edge cases, performance requirements, accessibility standards, and compliance considerations that vibe coding tends to gloss over. The demo working is not the same as the product being ready.

What Vibe Coding Is Good For

Despite the criticisms, vibe coding has real legitimate use cases:

Prototyping and validation. If you need to test whether an idea works before investing serious development time, vibe coding is the fastest path from idea to testable prototype. The quality does not matter yet — the question does.

Non-critical internal tools. A dashboard for your own use, a script that runs once a week, a utility that only you will maintain — these do not need the same standards as customer-facing production software.

Developers learning new domains. A backend developer who needs to build a frontend, a mobile developer learning web — vibe coding can produce a working result while you learn the patterns.

Experienced developers handling boilerplate. Senior developers using AI to generate the repetitive parts (CRUD operations, standard API integrations, test scaffolding) while writing the complex logic themselves. This is not really vibe coding — it is AI-augmented development — but it has become the norm.

What Comes After Vibe Coding: Agentic Engineering

Karpathy himself has since said that vibe coding is already passé. The term he now uses is agentic engineering.

The shift: AI has gotten capable enough that it is not just writing individual functions or components on request. It is running entire workflows autonomously — spinning up environments, writing code, running tests, fixing errors, iterating — with the developer acting as an overseer rather than a hands-on builder.

In agentic engineering, you define the goal, set the constraints, and review the output. The AI handles everything in between, including its own debugging and iteration cycles. Karpathy's framing: "You are not writing code 99% of the time. You are orchestrating agents who do, and acting as oversight."

Tools like Claude Code (the tool used to write this very post), Devin, and OpenHands are early implementations of this paradigm.

What This Means for Developers

Vibe coding will not replace experienced developers. What it will replace is a certain kind of junior developer work — writing standard components, implementing well-understood patterns, producing the boilerplate that senior developers find tedious. The demand for developers who can architect, review, debug, and make technical decisions will remain and grow.

The skill premium shifts. In a world where AI writes most of the code, the scarce skills are judgement (knowing what to build), architecture (designing systems that scale and maintain), code review (catching what AI gets wrong), and domain expertise (understanding what the output should actually do). Pure syntax knowledge depreciates. Everything else appreciates.

Speed advantage is real. A developer fluent in AI coding tools produces output significantly faster than one who is not. The productivity gap between AI-augmented and non-augmented developers is widening. This is the practical reason 92% of US developers have adopted AI coding tools — not ideology, but competitive necessity.

The bar for code quality review rises. If AI writes the first draft and humans review, the human's job becomes catching problems the AI creates — which requires a deeper understanding of what can go wrong, not just what the code does.

Vibe coding is a moment in the evolution of how software gets built — not the destination. What Karpathy named in 2025 has already been surpassed by the tools available in 2026. The trajectory is clear: AI writes more and more of the code; human developers provide more and more of the direction, judgement, and oversight.

Whether you call it vibe coding, agentic engineering, or something else that will be coined next year, the underlying shift is the same. Software development is becoming a higher-level activity — and the developers who adapt earliest will be the ones who define what it means to build software in the next decade.