Reckless vibe-coding is on the rise in tech companies and needs to stop

Many of my friends and I have been observing a rise in something the industry doesn't yet have an agreed-upon name as of writing. Some call it "meat proxying", some call it "AI psychosis", some call it "cognitive surrender", the list goes on. In this post I'll call it "reckless vibe-coding" just to make the concept easier to understand.

This is when a developer is assigned a task, be it to review a PR, work on a reasonably complex coding task, or even just asked a simple question, and all they do is prompt an agent "do X" and copy-paste the results back to you. They don't guide the agent toward the right path, they don't check if the agent's solution is appropriate, in fact they might not even look at the generated content at all. They will just prompt it and forward it to you regardless of what the output actually turns out to be. Despite being technically "prompted" by the human, in practice the work was 100% handled by the agent and the presence of the human made no difference whatsoever. This is referred to as "vibe-coding" within the industry, a term coined by Andrej Karpathy back in 2025. Oftentimes the person may fail to disclose, or, in more egregious cases, even actively attempt to hide the fact that the work was fully AI-generated.

If the task in question is to open a PR, the PR will then turn out to also have fully AI-generated titles/descriptions (often word-salads that are completely unreadable and unreviewable), and any further commits or review responses will also be fully AI-generated. Similarly, if the task is to review a PR, the person will just ask an agent to do it and post comments on the human's behalf. The human never sees the code in question or the comments that will be posted. Everything is done autonomously by the AI.

Regardless of the case, the pattern is the same: Despite the tasks and actions being in theory "signed" off by the human, in practice the work was handled entirely by AI with zero involvement of the human that prompted it besides the initial "do X" prompt. We've even been observing cases where the human attempted to use this to their advantage to strip themselves of any responsibility regarding the code they wrote and merged (e.g. "this code broke main and took down production? It was the AI who did this, not me!").

The difference between "regular" vibe-coding and what I call "reckless" in this case is that while the term vibe-coding is often used in the context of harmless prototypes and side-projects, here we're talking about code that is actually pushed to production and rolled out to real users.

This behavior is becoming increasingly prevalent in the software engineering industry and needs to stop. It should be considered completely unacceptable from an engineering culture point of view as it's no different from other unacceptable engineering concepts such as pushing untested code straight to main or performing similar kinds of reckless actions in production and/or involving company property. It should be treated with the same severity as these actions and punished equally.

It's easy to think of code reviews as the natural barrier against this kind of reckless code change, but the reality is that code reviews are no longer an effective measure against bad PRs, because the code reviews themselves are also largely delegated to AI nowadays. This means that untested AI-generated code can (and will) easily land in the main branch without a single human even being made aware of it, because every "human" you may spot having conversations and generating reviews can turn out to actually be an AI just pretending to be the human in question.

Even if a real human does spot the PR and notices that it was vibe-coded, the massive increase in PR throughput caused by the adoption of AI in tech companies has largely forced devs to either lower their quality bar or even abandon reviewing entirely to avoid being overwhelmed by review requests. This is a cultural problem where companies as a whole have to agree not to do this in the first place, much like the industry has already done in the past for things such as test coverage and PR etiquette.

This reckless form of vibe-coding is particularly worse than everything we’ve seen so far because it's a lose-lose:

• You get your time wasted and lose trust in your colleagues, because if all the vibe-coder does is proxy their work to an agent, you'd be better off ignoring that person entirely and talking to an agent yourself
• You get forced to do the vibe-coder's job for them, because you now have to not only spend time reviewing and testing agentic output that the vibe-coder did not, but likely also have to be the one dealing with the consequences of any nonsense code that they may have managed to merge when you were not paying attention
• The company's products suffer, as vibe-coded PRs are almost always complete nonsense (assuming that the vibe-coder did not attempt to guide the agent or check the resulting code in any meaningful capacity)
• The vibe-coder now has a ticking time bomb on their hands, as whatever gains in speed the vibe-coder may have gained by performing the reckless move are likely to be lost when the untested code in question inevitably results in a large-scale incident

It's essentially a novel way to silently scam both the company and the colleagues. I use the word “scam” on purpose here because it feels to me that some people have weaponized AI to trick their managers into believing that they are being productive by opening hundreds of PRs with bold claims of having fixed major issues, when in reality a deeper look reveals that this "work" is actually just a bunch of unverified vibe-coded nonsense that not only doesn't fix or improves anything at all, but also actively harms the products they are targeting and massively complicates the work of the devs who have to deal with the consequences of these reckless PRs and care about building good products and using AI responsibly.

Just because AI has made it very easy to generate code, it doesn’t mean that the generated code is guaranteed to be correct and safe. Even with “insert latest frontier model name here”, generating good code and code reviews still requires substantial human guidance and cannot be fully delegated to AI. Just like we as an industry have developed strict standards on what is and isn’t acceptable in the “old way” of writing code, we must now do the same in this new environment, starting with forbidding this form of reckless usage of AI in professional settings for anything that is not harmless prototyping.

(This post was originally published on Nov 08 2026 under the name “Being a meat proxy for AI should be grounds for getting people fired” with a more ranty / brain-dump format. I’ve since rewritten it to have a better structure and more neutral tone, but the message it communicates is largely the same as the original.)