What Is Vibe Coding in 2026? Agents, Review Bots and Where It Breaks
Last updated:
Vibe coding means describing what you want and letting AI write the code (Andrej Karpathy, X, 2025; Collins Word of the Year 2025). By September 2026 the tools run whole projects (Cursor Projects, September 10) and review their own pull requests (Cursor Security Review, September 23), but a human still checks the result. Reading code is still the skill.
Who coined vibe coding, and why is it the 2025 word of the year?
Andrej Karpathy, formerly of OpenAI and Tesla, posted on X on February 2, 2025: "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 his own workflow as seeing stuff, saying stuff, running stuff, and copy-pasting stuff, and later called the tweet "a shower of thoughts throwaway."
Nine months later, Collins named "vibe coding" its Word of the Year 2025, defining it as "the use of artificial intelligence prompted by natural language to write computer code." That is dictionary language for what Karpathy tweeted: talk to the model, run what it gives you, keep going.
What changed in 2026: from chat prompts to agents that run projects
Vibe coding in 2025 meant one chat window and one file at a time. By September 2026 the tools coordinate entire projects and grade their own output before you see it:
| Date | Feature | What it does | Who gets it |
|---|---|---|---|
| 2026-08-27 | Start from Scratch | Cloud agents work without a connected repo, with a live in-browser preview and one-click publish to Vercel | Paid plans, early beta (enterprise can opt out) |
| 2026-09-10 | Cursor Projects | A coordinator agent that "doesn't write code itself; it plans the work, delegates it to agents that implement it, and brings the finished work back to you to check," running for months on one project | Beta, rolling out to all users |
| 2026-09-23 | Rollouts and Security Review | Security Review scans every pull request for injection bugs, auth bypasses, and leaked credentials; Rollouts watches a deploy and "can also open a revert PR for review" | Teams and Enterprise plans only |
The moment the agent keeps the keyboard but you keep the review, the word changes too: what is agentic coding defines the loop these tools run and where your reading of the diff fits in it.
Do these tools mean you don't need to learn to code?
No. The job moved from typing code to judging it, and the vendors say so themselves. Cursor Projects hands you finished work "to check." Security Review lets you "dismiss one with a reason," which only works if you can tell a real finding from a false positive. Rollouts opens a revert PR "for review" and waits for a person to approve it. Each step assumes someone who can read the diff and judge it. You cannot approve what you cannot read.
That is a narrower question than "will AI replace developers," which our longer post on AI and junior developer jobs answers with survey data. This page's answer is about vibe coding specifically: the tools got bigger and faster, and the reading skill they assume matters more than it did.
When does vibe coding break? ("vibe coding hell")
Vibe coding hell is the 2 a.m. moment your users hit an error, Sentry lights up, and your own repo reads like a foreign language, because you never wrote it line by line. The AI did not understand your business rules. It pattern-matched them.
The triggers repeat across vibe-coded apps:
- Async bugs that look like random UI flicker
undefinedaccess that only fires on a slow network- Stale closures in React that make state lie to you
- Type coercion that breaks for one customer's locale and nobody else's
An agent that runs a multi-week project the way Cursor Projects does does not remove those bugs. It multiplies the amount of code you never personally wrote, so there is more of it to misread when something breaks. None of that is a prompting problem. A better prompt regenerates the same class of bug in a new shape. The fix is knowing enough JavaScript and React to read the stack trace yourself.
How do you fix vibe coding hell?
You backfill the fundamentals the agent skipped. Four steps, in order:
- Read every diff before you accept it. Treat an AI pull request the way Security Review treats one: something to check, not a black box to merge.
- Predict what the code does before you run it. If you cannot guess the output, you cannot debug it when the output is wrong.
- Learn to read a stack trace. Our guide to getting better at reading code is the drill for turning
Cannot read properties of undefinedinto a fix instead of a re-prompt. - Keep using AI, but practise without it too. How to keep learning to code without depending on AI has the method.
If step 3 is the gap, the free Learn to Code with AI course on Scrimba is built around that same loop: paste the AI's code, run it, find what it got wrong. In its debugging lesson, Guil Hernandez hands you a broken promo-code page with three bugs and leaves the last one for you to fix.
This is not about dropping AI tools. Developers who understand state and effects prompt better, review AI output faster, and stop the debugging loop instead of restarting it.
Vibe coding vs. tutorial hell: what's the difference?
Tutorial hell is passive consumption: you watch someone else type. Vibe coding hell is passive generation: you watch an AI type. Both skip the part where you struggle in short loops and build a prediction for what the code should do.
The exit is the same for both: smaller exercises, hand-written predictions, reading error messages without autocomplete open. If you recognize the watching version of this loop, how to escape tutorial hell covers the recall drills that work for vibe coders too.
Who is vibe coding actually for?
It fits prototyping, internal tools, personal projects, and experienced engineers who can audit AI output fast. It does not hold up alone in regulated systems, performance-critical paths you cannot profile, or teams without a review culture, and the 2026 team features assume that culture already exists: Security Review and Rollouts are built for a team that reviews pull requests, not a solo builder skipping that step.
If you are career-switching, do not mistake a shipped demo for a passed technical screen. A demo gets you noticed; you still need to explain the tradeoffs, the tests, and the error handling behind it in an interview. Our portfolio projects that get hired post is a concrete checklist for that gap, and how to learn React in 2026 covers the layer where vibe-coded apps break loudest.
Frequently Asked Questions
Vibe coding is describing what you want to an AI assistant, Claude, ChatGPT, Cursor, GitHub Copilot, and letting it generate the implementation. You steer by intent, not syntax. Collins Dictionary named it Word of the Year for 2025; the term traces to Andrej Karpathy's February 2025 post on X.
Andrej Karpathy, on X on February 2, 2025. He wrote that he 'fully gives in to the vibes' and later called the post a throwaway tweet. Collins Dictionary named 'vibe coding' its Word of the Year for 2025 nine months later.
Vibe coding hell is the moment an AI-generated codebase breaks in production and you cannot debug it because you never built a mental model of the logic. Common triggers: async bugs, undefined access, stale React state, and type coercion that only breaks for one locale.
Not by itself. It speeds up prototyping and lowers the barrier to shipping a first project. It becomes a problem when it replaces understanding, especially for debugging, security, and performance. Pair it with practice that forces you to type and predict code, not just accept completions.
Yes. Cursor Projects (September 10, 2026) delegates the writing to agents but still 'brings the finished work back to you to check.' Someone has to be able to read that work. The tools changed what you type; they did not change what you need to understand.
Learn to read the code your AI writes
Start free and build the debugging skill that agents and review bots still assume you have.
Sources and notes
- Andrej Karpathy, "vibe coding" post, X, February 2, 2025.
- Collins Dictionary, "Collins' Word of the Year 2025: AI meets authenticity as society shifts", announced November 6, 2025.
- Cursor, "Start from Scratch, Without a Repo", changelog, August 27, 2026.
- Cursor, "Projects", changelog, September 10, 2026.
- Cursor, "Rollouts and Security Review", changelog, September 23, 2026.
- Scrimba's Learn to Code with AI, reviewed on this site: 42 scrims opened, curriculum and lesson quotes verified against the course, September 2026.
If you want the next article in the series: how to learn coding without depending on AI.