Vibe coding means directing an AI agent to generate code from a plain-language description instead of writing it line by line. Whether that's a reasonable way to work or a real risk depends entirely on one thing: whether the output gets reviewed and verified before it ships.

The term has had an unusually fast rise and an equally fast reckoning. Coined by AI researcher Andrej Karpathy in 2025 to describe a specific, narrow practice, letting an agent write code while you barely look at it, "just for throwaway weekend projects," the phrase spread far past that original scope within a year. By 2026 it was being used for everything from quick prototypes to production systems, and the backlash arrived almost as fast as the hype.

What Vibe Coding Actually Meant, Originally

Karpathy's original definition was specific and narrow: you don't review the diffs, you don't dig into how the code works, and if something breaks, you prompt again rather than debug it yourself. He scoped this explicitly to low-stakes, throwaway work. Other developers drew a sharper line soon after: if an AI wrote every line but you reviewed, tested, and understood all of it, that's not vibe coding, that's using an AI agent as a fast typist. The distinction that got lost in the term's popularization is exactly the one that matters most: review, or no review.

Why the Backlash Happened

The concern that's built through 2026 isn't really about whether AI can write good code. It's about what happens when the "no review" version of vibe coding gets used somewhere it was never meant to. Security researchers have repeatedly found that a meaningful share of AI-generated code carries exploitable flaws when nobody checks it. Engineering leaders have started describing production incidents traced back to a component nobody on the team actually understood, because it was accepted straight from the agent. Even Karpathy has since distanced himself from unscoped use of the term, pointing instead toward what's been called "agentic engineering," a model with meaningfully more human review and judgment built back into the loop.

None of this means directing an AI agent is the problem. The problem is treating "the agent produced it" as equivalent to "it's correct," which was never true even in the term's original, narrow definition.

The Real Question Isn't Whether You Vibe Code, It's Whether You Verify

Almost every engineer now works this way to some degree: describing intent, letting an agent generate the implementation, and iterating from there. That's not inherently reckless. What separates a legitimate, fast way of working from the version causing production incidents is a single, checkable behavior: did the engineer actually read the code, test it, and confirm it does what was asked, or did they accept it on faith because it looked finished.

How HyperHat Measures This Directly

HyperHat scores a live, standardized 30-minute coding task completed alongside a real AI agent, across six dimensions: Task Decomposition, Prompt Quality, Verification, Iteration Efficiency, Recovery & Debugging, and Output Quality. Verification is scored explicitly, and if it isn't logged during the session, HyperHat withholds the score entirely rather than infer it from a finished-looking result. That rule exists precisely because of the gap the vibe coding backlash is about: a working demo and a verified result are not the same thing, and only one of them is safe to ship.

Frequently Asked Questions

What does vibe coding mean?
Vibe coding is directing an AI agent to write code from a plain-language description of what you want, rather than writing the implementation yourself. The term was coined to describe doing this without reviewing the output, originally scoped to low-stakes, throwaway projects.

Is vibe coding bad practice?
It depends entirely on what it's used for and whether the output is reviewed. Using it for a disposable prototype with no review is close to its original, narrow meaning. Using it for production code without reviewing, testing, or understanding what shipped is the practice driving the current backlash and the security concerns behind it.

Why did the term's inventor distance himself from it?
As the term spread to cover any AI-assisted coding, including production work with no review, it drifted far from Karpathy's original, narrow scope. He has since pointed toward "agentic engineering" instead, a model that puts more human oversight and judgment back into the process.

How can I tell if I'm vibe coding in a risky way?
The key question is whether you've actually read, tested, and understood the code an agent produced before shipping it. If the answer is no, and the code is going anywhere beyond a disposable prototype, that's the version of vibe coding associated with the current backlash.

How does HyperHat account for this?
By scoring Verification as one of six core dimensions and withholding the score entirely if verification isn't logged during the session. A finished-looking result isn't enough on its own; the assessment is built to confirm the work was actually checked.