What exactly is vibe coding?

Nowadays, code is generally no longer written line by line – countless frameworks, ready-made snippets, code completion and, of course, the first generation of AI code generators provide excellent support. Vibe coding is now breaking the classic pattern even further. The term was coined in early 2025 by AI pioneer Andrej Karpathy. The idea: a complete departure from the code base. Instead, an AI is conveniently told via language – spoken or written, of course – what functions the finished software or app should perform.

The AI then takes over the entire development process: it analyses the requirements, fills in any logical gaps that may have arisen in the task description, creates a rough architecture and functions, structures the code into modules and continuously adapts it based on user feedback. This cycle of generating, testing, observing and adapting makes the process very lightweight and therefore so interesting.

What added value does Vibe Coding offer and where are its limitations?

Thanks to its ease of use and accessibility, Vibe Coding opens up a wide range of applications. At first glance, however, the concept seems completely utopian and unrealistic in highly sensitive areas such as the automotive industry, where developers work with complex control units and safety-critical applications and are subject to strict standards and standardised processes. On closer inspection, however, a more nuanced picture emerges:

Possible applications and their benefits:

• Faster prototyping: For example, setting up a new screen for the infotainment system and testing various display elements.
• Generation of test cases: Automated derivation of test cases and their execution based on requirements – simply by means of a voice prompt.
• Support tools for process automation or data visualisation: Apps for process automation and integration across tools, as well as data visualisation or analysis tools that were previously written manually.

Difficulties in using Vibe Coding:

• Quality & maintainability: If you don’t really understand the code, it will be difficult later on to find bugs, improve the performance of individual sections, or implement changes that were not correctly captured by the AI.
• Compliance: ISO 26262, ASPICE, OEM audits – there are no shortcuts here. AI can provide support, but it cannot replace mandatory processes.
• Safety: AI can generate code that works, but contains potentially unsafe elements. This is particularly risky for safety-critical systems in cars. It is therefore essential to have people who check, test and understand the code. Without ‘human in the loop,’ it becomes dangerous.
• Hallucinations: Models sometimes generate code that looks plausible but may contain logical or safety-related errors.
• Lack of contextual depth: An AI model rarely knows all the project-specific requirements, dependencies, or design decisions.

Vibe Coding: Do we even need experts anymore?

The short answer: absolutely. Even though manual coding is becoming less common, in-depth knowledge remains essential. Developers need to classify the generated code, refine it and make it maintainable in the long term. There is no doubt that vibe coding can accelerate development processes, especially for early prototypes or standardised tasks. But as tempting as the advantages are, AI models do not always automatically deliver error-free, optimal code and have fundamental weaknesses.

The role of developers is therefore shifting: they are no longer ‘just’ authors of code, but increasingly also reviewers, architects and translators between requirements, AI systems and the code base. Ultimately, AI is only as good as the person using it.

Expert opinion: What our senior software engineer says

We speak with Thomas Schwiertz, Senior Software Engineer at Cognizant Mobility:

“Developers spend an estimated 30% of their time on direct coding. The majority of their time is spent on communication, coordination, compliance, code reviews and organisational processes – tasks that AI can only accelerate to a limited extent. Time savings achieved through AI therefore do not automatically translate into higher output, but often into refactoring, further training and quality assurance. “Human in the loop” remains central to responsibility, quality and security.”

Conclusion Vibe Coding: What market participants are saying now.

Vibe coding, large language models and agentic systems are changing the rules of the game – even in highly complex domains such as the automotive industry. However, no one today expects companies to develop software exclusively via voice input tomorrow. Nevertheless, impressive efficiency promises are already being communicated – often based on prototypes or laboratory scenarios that cannot yet fully reflect the complexity of real-world projects.

When it comes to new technologies such as vibe coding, Cognizant Mobility always focuses on the actual impact on the final software product. This allows us to provide clarity about the opportunities, risks and responsibilities that AI entails. True innovation does not come from hype, but from bold, well-considered steps that deliver measurable results. Vibe Coding still has to prove itself in this regard – even though the technological direction it has taken already appears more than promising today.