AI & digitalisation

Vibe engineering: how AI is redefining software development and what that means for your company

Moltomedia

Comparison of vibe coding and vibe engineering

*This graphic was partly created with the help of generative artificial intelligence and then edited manually

A good 15 months ago, Andrej Karpathy, co-founder of OpenAI and former Tesla AI Director, published a tweet that would turn the entire industry upside down: “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. […]” A few months later, Collins Dictionary named “vibe coding” its Word of the Year 2025. According to Exploding Topics, searches for the term shot up by 6,700 %. And suddenly it was no longer just developers discussing the topic, but CEOs, CTOs and boards.

But what is really behind it? And how does it differ from what reputable IT service providers should actually be delivering today?

🖥️ Vibe coding, vibe engineering, agentic engineering: three stages, not synonyms

Vibe coding in its original form means writing prompts, accepting code without reading the diffs and simply seeing what comes out. This works brilliantly for prototypes, internal demos or throwaway scripts. According to Business Insider, Meta product managers use it to build prototypes in hours and present them directly to Mark Zuckerberg. Klarna CEO Sebastian Siemiatkowski uses it to take conceptual work off his engineers’ hands.

The problem starts when such prototypes are supposed to go into production.

Simon Willison, one of the most influential software analysts in the scene, therefore proposed the term vibe engineering: a way of working in which AI tools are used intensively, but the human developer still makes the architecture decisions, carries out reviews and is responsible for quality assurance. Willison describes it as the honest middle ground: fast and AI-assisted, but not blind. 🤝

Karpathy himself followed up in spring 2026 and now talks about agentic engineering – the next step in development, in which specialised AI agents divide up the work of planning, coding, testing and iterating, while the senior engineer orchestrates instead of typing. According to ElevateX, those who work agentically still write code themselves only 5 to 15 percent of the time.

⚠️ What happens when vibe coding is used the wrong way

In 2025, Amazon ordered a 90-day reset of its code deployment controls after a series of incidents that were at least partly linked to its AI coding assistant Q. SVP Dave Treadwell spoke internally of “high blast radius changes”: software updates that spread uncontrollably because control mechanisms were missing.

A Gartner study warns: 75 % of technology decision-makers name governance as their primary concern when using agentic AI. 35 % see cybersecurity as the biggest obstacle to adoption.

⚙️ Vibe coding is not a mark of quality. It is a tool with clear limits on where it should be used. The only question now is: what does it look like when you take the idea further? This brings us to spec-driven & eval-driven development.

📐 From vibe coding to discipline

As a countermovement to unstructured vibe coding, a new paradigm has established itself since early 2025: spec-driven development (SDD).

➡️ Here, a machine-readable specification becomes the single source of truth – for code, tests and documentation alike – before a single line of code is written.

The Thoughtworks Technology Radar added SDD as one of the most important new practices in November 2025, and according to heise online, three levels of maturity can be distinguished:

  1. Spec-first: The specification is created once at the beginning.

  2. Spec-anchored: The specification is developed further during the process.

  3. Spec-as-source: The specification becomes the only artefact that humans still edit.

At the same time, eval-driven development (EDD) is gaining ground – an approach similar to the principle of test-driven development, but tailored to the particularities of AI systems. Because AI-generated code is probabilistic rather than deterministic, a simple pass/fail is no longer enough. With EDD, you therefore first define what “correct” actually looks like, including acceptance thresholds, before the actual code is written. Only then is development carried out and measured against these evaluation criteria.

For companies, this means: SDD and EDD give vibe engineering a concrete methodological foundation. Instead of relying on the gut feeling of a developer or an AI, you get a traceable process with clearly defined specifications and measurable success criteria – exactly what distinguishes production-ready software from a prototype.

💡 What you should take away from this

The question is no longer: “Should I use AI in software development?”

The question is: “Who makes sure it is used correctly, with the right specification and the right evaluation criteria?”

Because vibe engineering at production level needs people who:

  • understand architecture and system design

  • can review AI-generated code and take responsibility for it

  • write specifications precisely enough for AI agents to work from them reliably

  • define success criteria before development instead of checking afterwards whether something works

  • keep security, maintainability and scalability in view

  • choose the right tools for the right context

That is exactly the standard of an experienced IT agency: not prompting blindly, but combining engineering expertise with AI speed.

💙 At Moltomedia, we support companies at exactly this interface: from the first digital idea to a scalable solution in production. AI is a tool – not a substitute for responsibility.

How is your company already using AI in development, and where do you draw the line between experimenting and producing? We look forward to hearing about your experiences.

🔗 Sources: Andrej Karpathy, X (formerly Twitter), February 2025 | Collins Dictionary, Word of the Year 2025 | Simon Willison, simonwillison.net, October 2025 & May 2026 | MIT Technology Review, “From vibe coding to context engineering”, November 2025 | ElevateX, “Was ist Agentic Engineering?”, May 2026 | Gartner, Agentic AI Governance Report, 2025 | Business Insider, Meta Vibe Coding Report, 2025 | Exploding Topics, Search Trend Data, 2025 | heise online, “Fünf Tools für Spec-driven Development im Überblick”, 2026 | Eval-Driven Development 2026

Get the clarity you need to move your business forward now.

Start today with your free IT assessment and get a clear roadmap for your digital future in just a few days.

© 2026 Moltomedia. All rights reserved.