Vibe coding is supposed to keep you in charge of AI-written code — but does it actually, once you're in the weeds?

This explores where vibe coding fits between older autocomplete-style coding help and fully autonomous coding agents, and whether the human really stays in control. On paper, vibe coding sits in the middle. You aren't prompting function by function, and you aren't handing the whole job to an agent either. The design assumes you keep steering. The corpus suggests that assumption often fails in practice. Novices drift into the habits of passive agent users: they barely touch the code, test only the surface, and restart when things go wrong instead of fixing them Does vibe coding actually keep humans in the loop?. The tool keeps a seat for the human, but beginners often don't sit in it.

The classroom data shows what this looks like. Across 19 students, almost two-thirds of interactions were spent testing the running prototype. Only 7.4% touched the code, and 90% of those were reading, not editing Where do vibe coding students actually spend their debugging time?. So the human is still in the loop, just at a different level. They check behavior, not implementation. That changes what 'oversight' means. A person who only ever clicks through the app can catch visible bugs but not hidden ones.

One way to see what is distinctive about vibe coding is to ask who decides that something should be built. Moran contrasts it with generative UI, where the AI decides to create an interface element. In vibe coding the user starts the request, so the system is judged on whether it carried out the user's intent faithfully, not on whether it made good design choices Does who initiates a UI element matter for accountability?. That puts a lot of weight on the human stating intent clearly. Friedman argues that typed chat is a slow, tiring way to do this, and that canvases, visual intent maps, and refinement controls could take over part of the work Can task-oriented UI patterns reduce the friction of AI chat interfaces?. The tension doesn't go away, though. Generated interfaces are clearer but more rigid, while flexible ones push non-programmers back into thinking like engineers Do generated analysis UIs really work better than chat?.

Further along the spectrum, the question shifts from 'is the human paying attention?' to 'how does the system make itself visible to the human?' Proactive coding agents save effort but interrupt programmers' flow. Visible presence cues and shared context greatly reduce that disruption Do proactive AI coding agents actually disrupt programmer workflow?. Orchestration layers like Dr. Claw wrap an unchanged coding agent in persistent state and a traceable record of what it did, so the human can audit and recover the work instead of watching every step Can orchestration layers make coding agents more auditable?. Read together, these studies suggest that keeping a human in the loop works better as something built into the system's structure (logs, checkpoints, awareness signals) than as something you hope the user will keep doing.

The workflow also extends past the individual developer. At Microsoft, people started using coding agents mainly because their peers did, not because of seniority, and adopters merged about 24% more pull requests Do social networks drive adoption of new coding tools?. Evans warns that making building easier doesn't solve the harder problems: knowing which tasks are worth automating, and getting an organization to commit Does easier tool-building actually solve enterprise adoption problems?. The surprising takeaway is that vibe coding's weak spot isn't the AI's code. It's that the human role it depends on is fragile and often quietly abandoned. The answer seems to be designing systems that keep that role alive instead of assuming people will fill it.