What is vibe coding?
Vibe coding is a way of programming where you describe what you want in plain language and accept the AI’s code without reading it closely, judging only whether the result works. Andrej Karpathy coined the term in February 2025.
Where the term comes from
In a February 2025 post, Andrej Karpathy described giving in to the vibes and almost forgetting the code exists: he pasted error messages straight back in and accepted changes without reading the diffs. He called it vibe coding. The term spread quickly and is now used broadly for producing code with AI without reviewing it.
When it works
When the code is short-lived and mistakes are cheap, vibe coding is genuinely fast.
- A weekend project, a prototype or a throwaway script.
- Finding out quickly whether an idea holds up.
- Someone who doesn’t write code building a simple tool for themselves.
The risks
Code nobody read is code nobody understands. As the project grows, you won’t know where to look when something breaks.
- Security holes: endpoints without input checks, keys hard-coded in source, permissions that are too broad.
- Hidden bugs: logic that looks fine but fails on edge cases.
- Maintenance: duplicated, inconsistent code.
- Data loss: a migration or delete command nobody checked.
Safer vibe coding
You can cut the risk without giving up the speed. Keep everything in git and commit often; ask the agent to write and run tests; keep secrets in files outside the repo, such as .env; and have a person or a second model review changes before they go live. If the prototype is turning into a product, that’s the moment to switch to a supervised, agentic workflow.
Vibe coding in AgentVera
AgentVera doesn’t get in the way of quick experiments, but it adds a safety net for the moments you move on without reading the code.
- Checkpoints snapshot every turn, so a turn that broke things can be undone in one click.
- Automatic code review: when a turn ends, a second model reads the code and lists findings by severity.
- Verify shows whether your acceptance criteria actually passed, with evidence, instead of a bare “done” (Pro and Team).
- DB Guard scans migrations an agent wrote when its turn ends and flags risks such as DROP or an UPDATE without WHERE early.
- Checkpoints and session replayA snapshot of every turn; undo one in a click and replay a session step by step.
- Automatic code reviewWhen a turn ends, a second model reviews the agent’s code; findings link to file:line.
- AgentVera VerifyThe acceptance criteria in your message are checked against evidence by an independent verifier at turn end.
FAQ
Is vibe coding bad?
It depends on where you use it. For prototypes and personal tools it’s fast and fun. The problem is unread code going live in a product that holds user data.
What’s the difference between vibe coding and agentic coding?
In both, AI writes the code. In vibe coding the output is barely reviewed; in agentic coding the task is clearly defined, the change is reviewed and tests confirm it.
Which tools are used for vibe coding?
Coding agents such as Claude Code, Codex and Cursor, or tools that build apps in the browser. What matters more than the tool is how much you check the output.
Can I vibe code without knowing how to code?
For simple projects, yes. When something breaks or security is involved, someone who can read the code needs to look.