Reckless vibe-coding is quietly shipping to production

A MacBook with lines of code on its screen on a busy desk

Engineer Bruno Rocha has a name for something a growing number of developers are witnessing in professional codebases: reckless vibe-coding. It’s not the harmless prototyping that Andrej Karpathy described when he coined the term “vibe-coding” in 2025. It’s engineers prompting an AI agent with “do X,” copy-pasting whatever comes back, and shipping it to production without reading a single line.

What This Actually Looks Like

The pattern shows up across the full PR lifecycle. A developer submits a pull request with an AI-generated title and description that reads as word salad. Any follow-up commits and review responses are also fully AI-generated. The human never sees the code. In some cases, Rocha reports observing engineers actively hiding the fact that their output was entirely AI-generated.

Code reviews offer no protection either. When reviewers also delegate the entire review to an agent, untested AI-generated code can land in the main branch without a single human ever seeing it. Every “human” in the conversation thread can turn out to be an AI proxying for the person whose name is on the account.

Why This Is a Lose-Lose

  • You waste time and lose trust in colleagues. If a teammate is just proxying everything to an agent, you’d get more value talking to the agent yourself.
  • You absorb the cleanup work. You end up reviewing and testing agentic output the vibe-coder skipped, and dealing with any incidents from code they managed to merge unnoticed.
  • Products suffer. Rocha’s position is that vibe-coded PRs are almost always nonsense when the engineer made no meaningful attempt to guide the agent or verify the output.
  • The vibe-coder is sitting on a time bomb. Speed gains from skipping verification evaporate when the unreviewed code causes a production incident.

The Scam Framing

Rocha uses the word “scam” deliberately. His argument is that some engineers are opening hundreds of PRs with bold claims about fixing major issues, using AI throughput to look productive to managers, while the underlying work is unverified and actively harmful to the codebase. The massive increase in PR volume caused by AI tooling has also pushed many developers to lower their review quality bar just to keep up, which compounds the problem.

What Needs to Change

Rocha’s position is that this belongs in the same category as pushing untested code straight to main: a cultural violation that companies need to treat with the same seriousness as other established engineering standards. The fix is not a technical one. It’s a cultural agreement, similar to how the industry converged on test coverage expectations and PR etiquette, that fully delegating production work to an AI agent without any human verification is not acceptable.

The full post is worth reading if you’re thinking through AI usage policies for your team.

Stay on top of AI & Automation with BizStack Newsletter
BizStack  —  Entrepreneur’s Business Stack
Logo