What is Vibe Engineering?

The tech community is currently divided into two distinct camps. One side argues that AI is significantly accelerating software development, while the other worries that handing over entire tasks to AI is tantamount to 'outsourcing' a system that isn't yet reliable enough to build real-world software.

 

However, there's a major misunderstanding here. Building software with AI doesn't mean engineers stop doing anything. In fact, much of the work is still engineering, just at a different level. Instead of writing line after line of code, engineers now focus more on systems thinking, architectural design, and decision-making.

The AI ​​landscape is changing extremely rapidly.

To better understand "vibe engineering," we first need to look at the pace of AI tool development in recent times.

 

Cursor became a truly explosive AI-powered IDE in 2024, despite its launch in 2023. However, at that time, creating stable products using Cursor was still quite difficult. Even those familiar with the tool encountered many errors and had to make numerous adjustments.

Many people also remember Devin—a tool once touted as a 'junior AI engineer' capable of working independently. But in reality, Devin failed to complete projects as expected.

In just the last few months, things have changed rapidly. Spotify shared that their top programmers have almost stopped writing code manually since December. Anthropic's internal team is also said to have over 80% of its code implemented with AI support. Meanwhile, Andrej Karpathy argues that programming has changed more in the last two months than in previous years.

Another notable example is that Claude Opus 4.6 discovered 22 new vulnerabilities in Firefox in just two weeks, including 14 critical ones. This is equivalent to about one-fifth of the total number of critical bugs Mozilla will fix in all of 2025.

 

These advancements demonstrate that AI is improving rapidly. However, that doesn't mean engineering jobs have disappeared. Instead, the role of engineers is gradually changing.

You plan, the AI ​​writes the code.

If AI is already so powerful, why not just describe the requirements and let the AI ​​do everything?

The answer lies in the hardest part of software development: system planning and design. AI can write code very quickly, but it still needs human guidance.

AI should be viewed as an assistant rather than an architect. Users still need to think about the system structure, identifying components such as frontend, backend, security, and infrastructure before assigning tasks to the AI. With a clear overall picture, guiding the AI ​​becomes more effective.

 

In many cases, you can use an agent to research technology options beforehand, compare costs and performance, and then decide on a deployment strategy. If building a user authentication system, you can research services like Cognito or Auth0 to see which tool is more suitable.

When storing user data, you can have AI create a complete CRUD API, and then have another agent use that API in the main application. This approach significantly speeds things up, but still requires you to understand how the system works to avoid errors later.

If you don't define things clearly from the start, making adjustments later often takes much longer. This shows that even though AI helps speed things up, the build process is still genuine engineering work.

An engineer at Anthropic once shared an effective workflow: plan ahead, adjust the plan until it's sound, and only then let the AI ​​execute. When the plan is good enough, the code will usually be more stable as well.

Judgement and "Taste" are still very important.

One of the most irreplaceable elements is technical judgment and intuition. This is what helps you know when to ask questions, when to double-check, and when the results are unsatisfactory.

Those who have worked closely with software often know where system errors might occur, what needs checking, and which assumptions might be wrong. This also explains why many software engineers, despite their concerns about AI, are the ones best able to leverage these tools.

However, it's not just programmers who have an advantage. Product, design, or UX professionals can also leverage their experience to build products using AI. The key isn't knowing how to code, but understanding what constitutes a good product.

 

If you're new to this field, asking lots of questions will be very helpful. You can ask AI to assess security, analyze risks, or compare implementation options. Some people even use multiple different agents to compare results in order to reduce risk.

Judgment is the ability to know what to prioritize, while taste is your standard of quality. Both can improve over time as you work more with AI.

Which AI coding tool is leading the way?

For a time, Cursor held a prominent position among AI tools supporting programming. However, with the launch of their own tools by major companies like OpenAI, Anthropic, and Google, the competition has rapidly shifted.

Currently, Claude Code is the most talked-about name in the development community. This tool allows users to assign tasks to AI like a colleague, instead of just receiving code suggestions in the IDE as before.

Some people even say they almost never open their IDEs anymore, instead delegating tasks directly to the AI. Claude Code is also used for many purposes beyond programming, such as project management or task organization.

In addition to Claude Code, many people also combine it with the Codex as an additional support tool. Furthermore, Claude Code Skills is used to store project knowledge and guide AI to work more efficiently.

Bottlenecks still exist.

Despite rapid advancements, AI tools still have many limitations. One common problem is that AI can generate false information that still looks plausible.

In one instance, the AI ​​generated API documentation, which another agent then used to integrate the system. The build process was very fast, but debugging took a long time because the AI ​​had created a non-existent endpoint address.

Furthermore, serious incidents have occurred. An Amazon AI agent once deleted a production environment, causing disruptions lasting for hours. In another instance, Claude Code deleted a product database, and the Codex once wiped an entire user's hard drive.

 

Another problem is model drift, where the model's performance changes over time without the user's control. This has led the community to develop its own tools for monitoring AI performance.

Furthermore, one study showed that junior engineers using AI tend to have a reduced ability to understand code compared to those who write manually. This raises the question of whether programming skills can degrade over time.

AI isn't making software engineers redundant, but it's changing the way they work. The value of engineers is shifting from writing code to systems thinking, evaluation, and decision-making.

It could be said that software engineering is moving up to a higher level of abstraction, where AI handles the implementation while humans focus on direction and control.

Currently, AI is enabling much faster software development, but risks and limitations still exist. And at least for now, humans still play a central role in the product development process.