Our honest take
It has been a little over a year since Andrej Karpathy named the thing in February 2025: "fully give in to the vibes… forget the code even exists." It was half a joke, a Friday-night tweet about letting an AI drive while you barely glance at the diff. Then it escaped the lab. "Vibe coding" became a meme, then a workflow, then — somewhere in the back half of 2025 — an actual discipline that serious teams practice on purpose.
So a year on, the honest question is not "is vibe coding cool." It clearly is. The question is: what part of it was real, and what part was hype we should quietly stop repeating? We vibe code for a living and ship product with it every week, so we have opinions — but we also have scar tissue, which is the more useful thing to share.
What actually stuck
Three things from the original idea turned out to be durable, and they are the reason "vibe coding" outlived the meme phase.
Agents write most of the code now. This is not aspirational anymore. For a large share of everyday work — endpoints, tests, migrations, refactors, glue — a competent agent produces the first draft and often the final one. The human's fingers touch the keyboard far less than they did in 2023. That shift is genuine and it is not reversing.
English became a first-class input. The biggest practical change is that a clear written spec is now executable. You describe intent in prose, the agent turns it into a diff, and the quality of your prose is a real input to the quality of your software. People who write precise, structured English get dramatically better output than people who type "make it work." That is a new skill, and it is here to stay.
Verification became the job. This is the part that surprised everyone. When the machine writes the code, your value moves to deciding whether the code is correct. The teams that got good at vibe coding did not get good at typing fast — they got good at defining "done" and checking it ruthlessly. Verification discipline is the load-bearing wall of the whole practice.
What was hype
Now the uncomfortable half. The original framing had two ideas baked in that did not survive contact with production.
"Fully hands-off" was never true for real software. The demos where someone builds an app from a single sentence are real demos — of toys. The moment the thing has users, money, auth, or data, the hands come right back onto the wheel. Not to type, but to steer, review, and veto. Anyone selling you a fully autonomous path from idea to production is selling you the demo, not the job. The 2026 shift everyone noticed — the one Anthropic documented in its Agentic Coding Trends report — was precisely this: the interesting work moved from flashy demos to boring, durable workflow replacement, and that only happens when a human stays in the loop.
"Forget the code even exists" is the line that aged worst. You can forget the code exists right up until it breaks in a way the agent cannot see. Then you are debugging a system you never read, and the vibes evaporate. The developers who genuinely ship with AI have not forgotten the code — they have a working mental model of it and choose which parts to read closely. Forgetting is a luxury reserved for throwaway projects.
The word "vibe" undersells the craft
Here is our real gripe, and it is mostly with the name. "Vibe" makes it sound casual — like you lean back, feel it out, and code appears. The actual practice is the opposite of casual. Doing this well looks like: writing a tight spec, choosing the right model for the task, giving the agent exactly the context it needs and nothing it doesn't, running the change in isolation, verifying externally before you trust a single "tests pass," and keeping a checkpoint you can roll back to when it goes sideways at 2am. That is not vibes. That is engineering with a new kind of collaborator.
| The 2025 meme | The 2026 practice |
|---|---|
| Type a sentence, get an app | Write a spec, get a reviewable draft |
| Trust the agent's "done" | Verify externally, every time |
| One chat, forever | Split roles, fresh context, checkpoints |
| Forget the code exists | Choose what to read closely |
| Whatever model is trendy | Right model per task, swapped freely |
None of that makes it less powerful. If anything, the craft version is far more powerful than the meme version, because it scales past the toy stage. But calling it "vibe coding" invites people to skip the discipline and then act shocked when their AI-built app falls over. If you want the longer, more careful version of what the practice actually is, we wrote an honest guide to it.
So — future, or phase?
Our verdict, without hedging: it is the future, and it is durable, but the durable part is not the part the name advertises. The durable part is agents drafting most code, English becoming an engineering input, and verification becoming the core skill. The phase — the thing that will look silly in a few years — is the fantasy of the fully hands-off, don't-look-at-the-code, one-sentence-to-production developer. That was a demo aesthetic, not a working method.
We think in five years nobody will call it "vibe coding," the same way nobody calls using a compiler "vibe coding." It will just be how software gets built: humans specifying and verifying, machines drafting and iterating. The name will fade. The workflow will not.
The one genuinely open question underneath all of this is the one everyone actually means when they ask about the future — whether the humans stay in the loop at all, or get automated out of it. That deserves its own honest answer, and we gave it in will AI replace programmers. Short version: the job is changing far more than it is disappearing, and the people who treat vibe coding as a craft rather than a shortcut are the ones who come out ahead. We built The Vibe Father around exactly that bet — that the winners keep their hands on the wheel, they just stop doing all the typing.