On February 2, 2025, Andrej Karpathy — OpenAI co-founder, former Tesla AI director, one of the most respected names in machine learning — posted a tweet that the tech world is still arguing about:
"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."
He described how he barely touched his keyboard. He talked to his AI with voice. He hit "Accept All" without reading the diffs. When he got error messages, he copy-pasted them back into the chat and let the AI fix them. "The code grows beyond my usual comprehension," he wrote. "I'd have to really read through it for a while."
Within weeks, the phrase had exploded across developer communities worldwide. By March 2025, Merriam-Webster listed it as a "slang & trending" expression. By December 2025, Collins Dictionary named vibe coding its Word of the Year. By early 2026, 90% of developers worldwide were using AI coding tools daily, and 46% of all new code being written was AI-generated.
Then, almost exactly a year after coining the term, Karpathy declared vibe coding "passé."
So what is vibe coding, what did it actually change, where did it go wrong, and what is it evolving into? This guide gives you the complete, honest picture — from the original tweet to where the industry sits today in mid-2026.
What Is Vibe Coding? (Plain English)
Vibe coding is a style of software development where you describe what you want in plain English, let an AI generate the code, and iterate by testing the result rather than writing or reading the code yourself.
The name captures the spirit of the approach: instead of carefully specifying every line of logic, you work from a "vibe" — a general sense of what the outcome should feel like — and trust the AI to translate that intention into working code.
In Karpathy's own words, describing his workflow: "I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works."
The core workflow in practice:
- Describe what you want in natural language ("Build me a dashboard that shows real-time sales data with a chart")
- The AI writes the code
- You run it and see if it works
- If it doesn't, you describe the problem ("The chart isn't updating when I filter by date") or paste the error message back
- The AI fixes it
- Repeat until it works
That's it. No line-by-line authorship. No memorizing syntax. No deep dive into documentation. Just intent, output, and iteration.
What made this possible: by early 2025, AI coding models — especially Claude Sonnet, GPT-4, and the systems powering Cursor Composer — had become capable enough that for a wide range of common tasks, the first or second AI-generated attempt actually worked. The quality bar crossed some invisible threshold where "let the AI handle it" became a genuinely productive strategy rather than a frustrating exercise.
The Numbers That Show How Fast This Happened
<cite index="20-1">By 2026, 92% of US developers use AI coding tools daily, and 46% of all new code written by active developers now comes from AI.</cite> The vibe coding tools market is valued at $4.7 billion in 2026 and projected to reach $12.3 billion by 2027.
<cite index="16-1">A quarter of startups in Y Combinator's current cohort have codebases that are almost entirely AI-generated.</cite>
<cite index="22-1">41% of global code is now entirely AI-generated. In Java projects, that number spikes to 61%.</cite>
Cursor — the most popular AI-first code editor among professional developers — crossed $2 billion in annualised revenue by early 2026. Lovable, a browser-based builder for non-developers, scaled to $100 million in ARR within eight months of launch.
<cite index="16-1">Collins Dictionary named vibe coding its Word of the Year for 2025, and Merriam-Webster listed the term as "slang and trending" from March 2025.</cite>
These aren't projections. This is what happened in roughly 18 months from one tweet.
Who Is Vibe Coding For? (And Who It Isn't)
The honest answer is that vibe coding works differently depending on who's using it — and most articles present only one of these two realities.
For non-developers and beginners
This is where vibe coding is genuinely transformative. The barrier between having an idea and building a working version of it has collapsed in a way that simply wasn't true two years ago.
<cite index="20-1">A graphic designer with no programming experience built a children's app that turns emojis into stories for his 5-year-old.</cite> A marketing manager with no coding background can prototype a landing page, a data dashboard, or an internal tool in an afternoon. A student can build a portfolio project without getting stuck on environment setup and syntax errors.
For this audience, vibe coding is a genuine superpower. The limitation isn't vibe coding itself — it's knowing what good output looks like, catching problems the code has that you don't know to look for, and knowing when to stop trusting the AI's output on something safety-critical.
For experienced developers
The calculus is different. <cite index="20-1">Vibe coding lowers the floor. It does not lower the ceiling.</cite> A developer who understands code can review what the AI produces, catch security holes, and fix edge cases. A senior engineer using Cursor or Claude Code can achieve dramatically higher output — the AI handles boilerplate, routine implementation, and repetitive patterns while the developer focuses on architecture, judgment, and the genuinely hard parts.
<cite index="18-1">A developer who deeply understands system architecture can leverage a team of agents to achieve 10x or 100x productivity, while a novice will merely generate broken code faster.</cite>
For experienced developers, vibe coding is a productivity multiplier. It doesn't replace the expertise — it amplifies it.
The honest middle
There's a large group in between — developers with some experience who adopted vibe coding enthusiastically, used it for production systems, and ran into problems they weren't equipped to recognize or fix. This group is responsible for much of the backlash against vibe coding in late 2025 and early 2026, and for the wave of LinkedIn profiles jokingly renamed "Vibe Code Cleanup Specialists."
The Tools That Made Vibe Coding Possible
Several tools are central to the vibe coding ecosystem:
Cursor: The most popular AI-first code editor for professional developers. Works like VS Code but with deep AI integration — the model can read your entire codebase, propose multi-file edits, and iterate through errors automatically. Crossed $2B in ARR by early 2026.
GitHub Copilot: The most widely deployed AI coding assistant globally, with 90% of Fortune 100 companies using it. More conservative than Cursor — better at inline completion, less oriented toward full vibe-coding workflows.
Claude Code: Anthropic's terminal-based coding agent. Strong reasoning capability, particularly good at understanding complex codebases, debugging, and multi-step tasks. Works through your terminal rather than inside an IDE.
Lovable: Browser-based, designed for non-developers. Describe a web app, get a deployed version. Scaled to $100M ARR in eight months — until a security breach in 2025 became one of the key case studies for vibe coding's risks.
Bolt.new: Fast front-end prototyping from a single prompt. Best for quick UI mockups and landing pages.
Replit: Beginner-friendly platform with embedded AI agents. Accessible from a browser, good for students and quick experiments.
v0 (by Vercel): Generates React components from descriptions. Excellent for UI work without touching Tailwind or component libraries directly.
Each tool sits on a spectrum from "copilot assistance" (suggest as you type) to "full agent" (plan and execute multi-step tasks). The most powerful vibe coding workflows use the agent-style tools.
A Real Vibe Coding Workflow — Step by Step
Talking abstractly about "prompting an AI" doesn't convey what vibe coding actually feels like. Here's a realistic example:
The project: Build a simple link preview extractor — paste a URL, see the title, description, and OG image.
Without vibe coding: Set up a project, install Node.js, pick a framework, install dependencies, write a fetch function, handle CORS, parse HTML, extract meta tags, style the UI, handle errors. 3-6 hours for a developer who knows what they're doing. Weeks for a beginner.
With vibe coding in Cursor:
Prompt 1: "Build a single-page app where a user can paste any URL and see a preview showing the page title, meta description, and OG image. Make it look clean with a card layout. Use vanilla JS, no framework."
The AI generates an HTML file, a JS file, and a CSS file. You open it. The fetch works but the OG image isn't showing.
Prompt 2: "The OG image isn't loading. The console shows a CORS error."
The AI rewrites the fetch to use a CORS proxy. Now the image loads. The card looks a bit plain.
Prompt 3: "Make the card look more like a Slack link preview — left border, favicon, cleaner typography."
The AI updates the CSS. Now it looks good.
Total time: 20 minutes. Total code written by the developer: zero lines.
This is the experience Karpathy was describing. For a throwaway tool, a demo, a weekend project — it's remarkable. Our own Link Preview Extractor does exactly this — try it to see the kind of utility that vibe coding makes trivially fast to prototype.
Where Vibe Coding Goes Wrong — The Honest Risks
Most enthusiastic vibe coding articles stop before this section. They shouldn't, because these risks are real and documented.
Security vulnerabilities at scale
<cite index="20-1">45% of AI-generated code contains security vulnerabilities such as command injection and hardcoded API keys.</cite>
This isn't a minor footnote. When developers "Accept All" without reading diffs — as Karpathy described in his original post — they're accepting code they haven't reviewed. In a throwaway project, that's fine. In a production system handling user data or financial transactions, it's a serious risk.
The Lovable breach: Lovable, one of the most successful vibe coding platforms, experienced a significant security incident in 2025 that became a case study in the risks of deploying unreviewed AI-generated code at scale.
The Replit database deletion: An AI agent in Replit deleted a user's database during an automated workflow — a dramatic example of what happens when AI takes irreversible actions without human checkpoint.
These weren't edge cases. They were high-profile incidents at major platforms, in the same year those platforms were scaling explosively.
Technical debt without understanding
<cite index="16-1">Some developers commit AI-generated code without comprehending its functionality, leading to undetected bugs, errors, or security vulnerabilities.</cite>
Code that works today may be unmaintainable tomorrow — especially when the developer who "wrote" it (by accepting AI output) has no mental model of how it actually works. Six months later, when a bug appears or a requirement changes, debugging code you've never read is significantly harder than debugging code you wrote.
The trust paradox
<cite index="23-1">Usage was universal by 2026, but trust in AI code had dropped from 40% to 29% in twelve months.</cite> Almost everyone is using it; fewer people trust it unconditionally than a year ago. This is a healthy maturation — the early uncritical enthusiasm is being replaced by a more calibrated understanding of where AI code is reliable and where it needs scrutiny.
The "Vibe Code Cleanup Specialist" problem
<cite index="19-1">For many developers, AI coding agents often actually equal more work, and that's measurable in how the most experienced people in the room are spending their hours.</cite>
The hidden cost of vibe coding at scale is that experienced developers spend increasing time reviewing, fixing, and explaining AI-generated code to the team — work that's less visible than shipping features but just as necessary.
What Most Articles About Vibe Coding Don't Explain
It already evolved — Karpathy himself moved on
<cite index="15-1">Exactly one year after coining the term, Karpathy declared vibe coding "passé," introducing a more structured paradigm he calls "agentic engineering."</cite>
The distinction matters. Vibe coding in its original form — "Accept All, forget the code exists, paste error messages and hope" — was explicitly framed as appropriate for throwaway weekend projects. It was never intended as a professional development methodology.
<cite index="18-1">Agentic engineering is the mature successor: "agentic" because the default is that you are not writing the code directly 99% of the time — you are orchestrating agents who do — and "engineering" to emphasize that there is an art, science, and expertise to it.</cite>
The difference in practice:
- Vibe coding: you write a prompt, accept the output, hope it works
- Agentic engineering: you design the architecture, define the constraints and quality requirements, assign tasks to AI agents, review outputs critically, and own the result
One is a fun experiment. The other is a professional methodology.
The role of MCP in making agentic engineering work
If vibe coding is "tell the AI what to build," agentic engineering is "give the AI the tools to actually build it properly." That's where protocols like MCP (Model Context Protocol) become central infrastructure.
When an AI coding agent can connect directly to your GitHub repository, your test runner, your database, your deployment pipeline — through standardized MCP connections — it can do far more than generate code snippets. It can run tests, check outputs, iterate on failures, and deploy results. That's the technical foundation of agentic engineering.
Our complete guide to Model Context Protocol (MCP) explains this infrastructure layer in detail — if vibe coding and agentic engineering interest you, understanding MCP is the natural next step. You can also see how AI coding agents are reshaping developer productivity and our roundup of the best AI coding agents in 2026 for the current tool landscape.
The New Developer Identity — From Coder to Orchestrator
<cite index="22-1">The winners in 2026 won't be the fastest typists or the deepest syntax experts. They will be Orchestrators: the people who have realized that in an age of infinite code, the only bottleneck left is human intent. Vibe coding is the transition from Instruction to Intention.</cite>
This framing — the "Orchestrator" — captures the emerging developer identity better than either "vibe coder" or "traditional programmer."
An Orchestrator:
- Understands what good architecture looks like (even if AI writes the implementation)
- Knows how to break complex problems into tasks that AI agents can execute reliably
- Can review AI-generated code critically — catching security issues, logic errors, and maintainability problems
- Sets quality standards and test coverage requirements before AI starts writing
- Understands the domain deeply enough to know when AI output is plausible-but-wrong
<cite index="21-1">The value shifts upward: from syntax and implementation toward judgment, taste, and oversight.</cite>
This is a more strategic role, not a less important one. The Orchestrator isn't replaced by AI — they're multiplied by it. The bottleneck moves from "can you write the code?" to "do you know what the right code should do?"
What Vibe Coding Means for Indian Developers
India has one of the world's largest developer populations, and vibe coding is reshaping the landscape in ways specific to the Indian tech ecosystem.
The fresher opportunity: entry-level developers who adopt AI-assisted workflows early — learning to prompt effectively, review AI output critically, and work at the agentic layer — will have a significant productivity advantage over those who don't. The question isn't whether to use AI coding tools; it's how deeply to understand what they produce.
The upskilling imperative: the skills that remain most valuable — system design, security awareness, code review, understanding business requirements — are exactly the skills that traditional computer science education emphasizes. These become more important, not less, as AI handles more implementation.
The IT services angle: India's large IT services industry (TCS, Infosys, Wipro, HCL and others) is deeply involved in enterprise software maintenance and development. Agentic engineering — AI agents that can handle routine implementation while human engineers handle architecture and review — is directly applicable to these workflows. The developers who learn to supervise AI agents effectively will be the most productive team members.
The freelancer opportunity: Indian freelancers on platforms like Upwork and Toptal can now take on projects that would previously have required larger teams — an AI-assisted solo developer can produce work at a velocity that was previously only possible with a team. This is reshaping what's possible for independent developers.
For students and self-taught developers: the most practical starting point is to pick one AI coding tool (Cursor or Claude Code are good starting points) and use it for a real project — not just toy exercises, but something you actually want to build. Learn where it's reliable, where it needs your oversight, and how to communicate requirements clearly enough that the output is useful.
Should You Learn Vibe Coding in 2026?
The honest answer by audience:
If you're a non-developer with an idea: yes, absolutely. The tools are good enough that you can build real things. Start with Lovable or Bolt.new for web apps, or Replit for smaller projects. Be aware that you'll need to learn enough to validate what the AI produces before deploying anything with user data or payments.
If you're a student learning to code: use AI tools, but don't let them replace understanding. Use vibe coding to build things faster, but make sure you can read and explain the code the AI writes. The fundamentals — data structures, algorithms, system design — still matter, because they're what you'll use to evaluate AI output.
If you're an experienced developer: you're probably already using these tools. The question is whether you're using them at the copilot level (inline suggestions) or the agentic level (full task delegation). Moving toward agentic workflows — with proper review practices — is where the biggest productivity gains are.
If you're a tech manager or CTO: vibe coding is already happening on your team whether you've sanctioned it or not. The 90% adoption number means your developers are using these tools. The productive response is to establish review practices, security requirements, and testing standards for AI-generated code — not to prohibit tools that aren't going away.
Common Misconceptions About Vibe Coding
Misconception 1: "Vibe coding means coding is dead" Coding isn't dead — the nature of the skill is shifting. Writing every line manually is becoming less central; understanding architecture, reviewing output, and setting quality requirements is becoming more central. The role evolves; it doesn't disappear.
Misconception 2: "Anyone can now build production software without coding knowledge" You can prototype. You can build tools for personal use. For production systems with real users, real data, and real security requirements — you need either coding knowledge to review the AI's output, or a developer who can. The barrier to prototyping has collapsed; the barrier to safe production deployment has not.
Misconception 3: "Vibe coding is what Karpathy currently recommends" Karpathy himself moved on from vibe coding to agentic engineering in early 2026. The original "Accept All, forget the code" framing was always described as appropriate for throwaway projects — not professional development. His current recommendation is disciplined human oversight of AI agents, not blind acceptance.
Misconception 4: "AI coding tools are just autocomplete" The current generation — Cursor, Claude Code, GitHub Copilot Workspace — operate at the agent level: planning multi-step tasks, using tools (file system, terminal, browser), running tests, iterating on failures, and making changes across multiple files simultaneously. This is categorically different from the autocomplete Copilot of 2022.
The Practical Starting Point
If you want to actually experience vibe coding rather than just read about it, here's the simplest real starting point:
- Install Cursor (free tier available) or open Claude Code
- Think of one small tool you actually want — something useful for your daily work
- Describe it in a few sentences and let the AI build a first version
- Run it. Note what's wrong or missing.
- Describe the problems and let the AI fix them
- Read through the final code before using it for anything important
That last step is what separates responsible vibe coding from the kind that causes problems. Understand what you're running — even if you didn't write it.
For tools that support your own development workflow while you build: our JSON Formatter is essential for inspecting API responses, our Base64 Encoder handles authentication payloads, our UUID Generator creates IDs for testing, and our Text Difference Checker is useful for comparing AI-generated code versions before accepting changes. Our Password Generator is handy for creating test credentials — all free, all in your browser, no signup.
For a deeper understanding of the infrastructure that powers agentic engineering workflows — the layer that sits below vibe coding tools — our AI autonomous agents explainer and multi-agent AI systems guide cover how these systems actually work.
Summary
Vibe coding started as a tweet, became a cultural moment, and is now settling into something more mature and more honest.
The original version — full surrender to the AI, no code review, paste error messages and hope — was always a prototype methodology dressed up in provocative language. It was fun, it was useful for throwaway projects, and it made a genuine point about how good AI coding had gotten.
What it's becoming — agentic engineering, orchestrated workflows, AI agents with human oversight — is more rigorous, more scalable, and more appropriate for professional software development. The developer's role doesn't disappear; it moves up the stack, from implementation to architecture to judgment.
The numbers make the direction clear: 90% adoption, $4.7B market, Word of the Year. This isn't a passing trend. But the mature version of this trend looks less like "forget the code exists" and more like "understand the code well enough to trust it."
That's probably a good thing.
Building with AI agents and want to understand the infrastructure layer? Our guides on Model Context Protocol (MCP), best AI coding agents, and how AI code assistants are reshaping developer productivity complete the picture. And explore our free Developer Tools for the utilities that speed up your AI-assisted development workflow.
