What is vibe coding?

Over the past year, people working with software development and AI have been sharing their experiences online, experiences that show a shift in how code is written. Then, in early 2025, a tweet by former OpenAI founding member Andrej Karpathy gave this practice a name: vibe coding.

What is vibe coding?

At its core, vibe coding is the process of building applications using natural language commands rather than writing code manually. Developers – or in some cases, non-technical users – interact with AI agents through text or voice, describing what they want the software to do. The agent then interprets these instructions and writes the code on their behalf. In some cases, the agent also tests and debugs the application, making the time between idea and execution shorter.

Replit, a collaborative coding platform with an integrated AI developer experience, has been a supporter of this trend. It reports that the volume of development tasks performed entirely by AI has been doubling approximately every seven months. And, according to the company, its AI-powered agents are now capable of writing complete features or entire applications, based solely on user prompts.

AI does the heavy lifting

This process is not an extension of autocomplete tools like GitHub Copilot, but rather it is a shift in the developer's role – from someone who writes code line-by-line, to someone who directs and critiques an AI that is doing much of the heavy lifting. Replit's agents, for example, operate within a fully integrated development environment (IDE), and can carry out tasks such as setting up databases, building UIs, connecting APIs, and debugging errors, all from a conversational prompt.

This way of working is already having an impact in early-stage startup environments. Y Combinator has noted that a growing number of companies in its accelerator cohorts are relying heavily on AI-generated code. In some cases, startups report that 95% of their codebase has been written by AI, enabling them to move faster and build more ambitious products with smaller teams.

The challenges with vibe coding

As much as vibe coding can achieve, it also introduces new challenges – because as the amount of auto-generated code increases, questions are raised around, for example, software quality, maintainability, and security. While AI can generate syntactically correct code, it may not always account for long-term maintainability, best practices, or edge cases (rare or unexpected situations that fall outside the normal range of inputs or conditions, for example, users entering blank fields, unexpected sensor data, missing files, or failed network connections). Although these scenarios may be uncommon, they can cause serious failures if not properly handled.

Nonetheless, vibe coding is showing a change in how we think about programming, and so by making the act of writing software more accessible, there is potential to lower barriers to entry for new developers, while also changing the expectations and workflows of experienced ones. If embedded system interfaces, hardware control scripts, and test environments can be generated by describing the problem rather than manually crafting the code, the design process could become faster and more exploratory.

Sending out the future vibes

Vibe coding is a marked step forward in what narrow AI can achieve. These systems are very capable within specific tasks – like interpreting natural language, generating functional code, and responding to iterative feedback – but they still rely on human input, context, and judgement. However, it is an example of how AI systems are beginning to perform more creative and open-ended tasks, and acting less like tools and more like collaborators.

In this sense, vibe coding could be seen as a precursor technology – one of the many, many steps – towards artificial general intelligence because as developers grow used to working alongside AI and these systems become more capable, the crossover between human intention and machine execution is likely to continue changing, and vibe coding may well become a standard part of the engineering toolkit – not just a trend in software development, but a transformation in how we build technology.