Vibe Coding Shows How AI Is Rewriting Who Gets to Build Software

What vibe coding actually is

For most of computing history, building a website or an app required knowing how to write code. That barrier is eroding. Under a practice now widely called "vibe coding," a person describes in ordinary English what a program should do, and an AI agent tries to build it 1. Harvard's framing is careful. The output may have plenty of limitations, but it will still exceed what someone without substantial technical skills could have produced alone 1.

That comparison is the key to the trend. The measure is not whether a vibe-coded app matches the work of an experienced engineer. It is whether a non-programmer can now ship something that previously would have been out of reach. On that measure, the answer appears to be yes, and Harvard presents the practice as a possible preview of how people will work alongside AI more broadly 1.

The tooling is moving fast

The vendor landscape shows how quickly the idea is becoming a product category. One AI news feed devoted to vibe coding lists a steady run of releases built around the same premise of plain-language instructions turned into working software 2.

Several of those releases show where the practice is heading:

• Krea has added custom apps to its creative agent. Users can generate bespoke image and video tools from a single prompt 2. This moves vibe coding beyond websites and into tools for creative work.
• LM Studio's Bionic agent now includes a shared Excalidraw canvas. Developers can sketch diagrams, wireframes and system designs that the agent can read and edit with them 2. The interface is no longer limited to text, and a rough drawing can serve as part of the specification.
• Lovable has added a model called Opus 5.5 and says it now builds apps in 57% fewer steps 2. That figure comes from the company, but it reflects how competition in this space centers on speed and fewer back-and-forth rounds.

Taken together, these updates suggest the category is splitting in two directions. Tools like Krea and Lovable aim at people who may never read the code underneath. LM Studio's canvas approach seems aimed at developers who want a faster, more visual way to collaborate with an agent.

The security gap

The same feed that tracks these launches also carries a warning. A deep-dive on reviewing vibe-coded apps before deployment cites an API key theft involving METR. It argues that authentication, credential access and permissions need to be checked before anything goes live 2.

This is the issue behind Harvard's remark about "plenty of limitations" 1. A program can appear to work in a demo while mishandling secrets or granting too much access. A person who cannot read the code may not be able to spot those flaws. The easier it is to produce software without technical knowledge, the more likely it becomes that software gets deployed by people who cannot judge whether it is safe.

Where the sources agree, and where they differ

Both sources treat vibe coding as real and consequential. They differ in emphasis.

• Harvard looks at the bigger picture. It treats the practice as a lens on how AI may change who gets to create and what expertise means 1.
• The industry feed looks at the ground level. It covers model upgrades, interface experiments and step-count reductions, along with a reminder that security review cannot be skipped 2.

The two views fit together. The broad story is about wider access to building software. The practical story shows that access arriving through a stream of product updates, each making the process faster or more flexible.

My reading

Vibe coding seems best understood as a shift in where human effort goes, not as the end of programming. Describing what you want becomes the main skill. Checking what you got becomes the main responsibility. Vendor improvements like fewer steps, visual canvases and one-prompt apps all reduce the effort of describing. Very little in the current wave of announcements addresses the work of checking.

That imbalance is the part worth watching. If Harvard is right that vibe coding previews a wider AI future 1, that future may favor people who can make things quickly. The people who can confirm those things are sound may be scarcer. The security advice circulating alongside the product news 2 suggests that review skills could become the most valuable form of technical literacy in a world where almost anyone can generate code.