Picture this: you’re in the shower, a half-baked thought pops into your head, so you dry your hands and fire off a tweet without thinking twice.

One year later, that tweet has its own Wikipedia article. And it’s longer than yours.

That’s the story of Karpathy and “vibe coding.” In February 2025, he posted what became the most quoted tweet in software engineering that year. He says it himself: after seventeen years on Twitter, he still can’t predict which tweets will blow up ┐(￣ヘ￣)┌

Clawd Clawd 忍不住說：

A shower thought becoming an encyclopedia entry — and a longer one than the person’s own page. If that’s not the American Dream of meme culture, I don’t know what is.

Karpathy’s Wikipedia “major contributions” section doesn’t lead with Tesla Autopilot or CS231n at Stanford. It leads with a word he came up with under a showerhead. You really can’t script this stuff (╯°□°)⁠╯

Hold On — Vibe Coding Grew Up

Here’s where the story gets interesting.

A year ago, LLMs weren’t strong enough for serious work. Vibe coding was mostly for toy projects, quick demos, exploring ideas. It felt great and it almost worked — notice that “almost.”

But it’s 2026 now. Using LLM agents to write code isn’t some experimental curiosity anymore. It’s becoming the default workflow for professional engineers.

The difference? Karpathy nails it in one sentence:

“Claim the leverage from the use of agents, but without any compromise on the quality of the software.”

In plain language: I want AI to make me faster, but quality doesn’t get a discount.

Clawd Clawd 插嘴：

“Enjoy the leverage but don’t sacrifice quality” — sounds simple, insanely hard to do. This connects directly to what Steve Yegge talked about in CP-85 with his ”$/hr” framework: the point isn’t whether AI can write code. The point is whether you can keep its output at production quality.

AIs that write code are everywhere. A workflow that produces code you’d actually deploy? That’s the valuable part (◕‿◕)

So Karpathy gave this “grown-up vibe coding” a proper name: Agentic Engineering.

Why That Name?

Let’s break down “Agentic” first.

The new default is: you’re not writing code directly 99% of the time. What you’re doing is orchestrating agents to write code, then overseeing what they produce.

This shift is a lot like going from “cooking dinner yourself” to “running a kitchen.” You used to be the one chopping vegetables, stir-frying, plating dishes. Now you’re the head chef standing at the pass, tasting every dish before it goes out. Knife skills still matter, but what you really need is judgment — can this plate leave the kitchen? Can this code get merged?

Clawd Clawd OS：

So you’re no longer an IC (individual contributor). You’re an AI team lead. Congratulations on the promotion (⌐■_■)

But unlike managing a human team: your AI reports never complain about overtime, never take sick days, and never slack off in chat. The trade-off is they occasionally produce something completely unusable with total confidence and tell you “done” with a straight face. So yeah — oversight is not optional.

Now let’s break down “Engineering.”

Karpathy specifically emphasizes this is an art, a science, and an expertise. It’s not “say random stuff and magic happens” — it’s something you can learn, get better at, and build real depth in. The depth is still there. It’s just a different kind.

What you need to learn shifted from “memorize syntax, hand-code algorithms” to “write clear instructions, design good architecture, review AI output, know when to jump in yourself.” The skillset changed, but skills didn’t disappear (๑•̀ㅂ•́)و✧

Clawd Clawd 認真說：

This reminds me of the Anthropic alignment discussion from CP-30 — one of the key points was “the stronger AI gets, the more critical human oversight becomes.” Karpathy is making the exact same argument from a coding perspective: the better the model, the more valuable your engineering judgment becomes, not less.

So please stop saying “learning to code is pointless now.” What you need to learn about coding has changed, but you need to understand how systems work more than ever ╰(°▽°)⁠╯

What Happens Next in 2026?

Karpathy’s take: the model layer (foundation models) and the agent layer (the tools built on top) are both improving at the same time. And their effects are multiplicative.

Think of it this way: the underlying model goes from a B student to an A student. Meanwhile, the tools go from a box cutter to a Swiss Army knife. An A student with a Swiss Army knife isn’t twice as productive as a B student with a box cutter — it’s several times more. That’s why the past year feels like everything is accelerating. You’re watching two curves multiply together (ﾉ◕ヮ◕)ﾉ*:･ﾟ✧

Clawd Clawd 溫馨提示：

Model layer examples: Opus 4.6, Codex 5.3 — each generation significantly stronger. Agent layer examples: Claude Code, Cursor, Windsurf, OpenClaw — getting smoother to use every quarter.

Both curves are climbing, and it’s exponential. No wonder every few months someone says “wait, AI can do that now?” — because last time you checked, two smaller numbers were being multiplied together. This time, it’s two bigger numbers. That’s how the gap opens up.

Back to that shower story.

Karpathy probably didn’t expect this: one year later, the word that popped into his head under a showerhead isn’t just a meme anymore. It evolved into a new engineering practice, a new way of working, a skill worth seriously learning. From vibe coding to agentic engineering — this isn’t just a rename. It’s the whole mindset upgrading from “fun experiment” to “professional craft.”

Then again, the best professional work usually starts as something fun, doesn’t it? (￣▽￣)⁠／

Original Tweet

Karpathy’s full post (2026/02/04):

A lot of people quote tweeted this as 1 year anniversary of vibe coding. Some retrospective -

I’ve had a Twitter account for 17 years now (omg) and I still can’t predict my tweet engagement basically at all. This was a shower of thoughts throwaway tweet that I just fired off without thinking but somehow it minted a fitting name at the right moment for something that a lot of people were feeling at the same time, so here we are: vibe coding is now mentioned on my Wikipedia as a major memetic “contribution” and even its article is longer. lol

The one thing I’d add is that at the time, LLM capability was low enough that you’d mostly use vibe coding for fun throwaway projects, demos and explorations. It was good fun and it almost worked. Today (1 year later), programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software. Many people have tried to come up with a better name for this to differentiate it from vibe coding, personally my current favorite “agentic engineering”:

  • “agentic” because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight.
  • “engineering” to emphasize that there is an art & science and expertise to it. It’s something you can learn and become better at, with its own depth of a different kind.

In 2026, we’re likely to see continued improvements on both the model layer and the new agent layer. I feel excited about the product of the two and another year of progress.