Why agentic engineering changes everything

Anyone who believes that AI-supported development only means faster coding has not understood the game. The real challenge is not in the prompt, but in the engineering: how do you create environments in which agents reliably produce high-quality code? Cleanup crews that refactor automatically, retrospectives with agents and production lines instead of individual features are the new tools - but only if the architecture is right. Because the Dora study shows brutally clearly: with bad principles, AI only makes you worse faster.

Podcast Episode: Why agentic engineering changes everything

In this episode, I talk to Benedikt Stemmildt about agentic engineering - and why “vibecoding” is the wrong term for what’s really happening right now. Benedikt shows that it’s not about faster coding, but about architecture, quality assurance and creating environments where AI agents produce good code. We talk about why cleanup crews are more important than perfect prompts, how retrospectives work with agents and why teams suddenly step on each other’s toes when everyone is working agentically at the same time.

“If you have bad principles and use AI, you’ll get worse faster. And if you have good principles, you get better faster.” - Benedikt Stemmildt

As a technologist since childhood, Benedikt Stemmildt has been dedicated to improving the working environment of developers for over 20 years. Developer Experience is his passion: his mission is to support teams in adapting the new way of working of “Agentic Software Engineering” and using it profitably.

Highlights der Episode

• Agentic engineering replaces vibe coding: it is architecture work, no more spontaneous code generation.
• AI makes you worse faster with bad principles - better faster with good ones.
• Code quality is created by cleanup agents that continuously deduplicate and refactor.
• Teams of five developers with agents stand on each other’s feet - smaller cross-functional units work.
• Throw away configurations after model updates: New defaults can be better than old skills.

Agentic Engineering: How AI is changing software development and quality

From “vibe coding” to agentic engineering

In the world of software development, new terms are constantly clashing with old principles. New things emerge, old things remain useful, many things are tried out. A good example of this is the change from “vibe coding” to “agentic engineering”. While “vibe coding” was understood more as loose creative interaction with generative AI, “agentic engineering” brings engineering thinking back into the discussion. The focus is shifting: it is no longer just about generating code, but about the entire process behind it - with quality, architecture and systematics.

The importance of principles - good faster or bad faster?

When AI-supported tools are used in software development, this does not happen in a vacuum. The old principles such as code quality, clean architecture and well-chosen teams still count. As Benedikt Stemmildt emphasizes in the podcast, the use of AI reinforces the effects of these principles: Those who already have bad practices will produce even worse code faster with AI. Those who live by good principles, on the other hand, get a boost from AI. In short: AI accelerates the direction in which a team is already heading.

Clean-up agents: quality emerges after the first throw

One exciting aspect of agentic engineering is the “cleanup” principle. The first draft does not have to be perfect - just like in agile methods, it is the rework that decides. Here, development is supported by AI agents that automatically find errors, eliminate redundancies and generate tests. Instead of human clean code engineers, specialized agents take on the task of keeping the code base lean and clear. This gives developers more freedom for experimentation and creativity.

Models, tools and the right way to handle new features

The discussion about the latest AI model is hot, but not everything revolves around changes and upgrades. Benedikt Stemmildt recommends starting with blank settings for new models to familiarize yourself with their behaviour instead of dragging along old, possibly obstructive habits. Agentic engineering also offers the possibility of building software development lines: A process chain in which different AI models are used specifically for different tasks - for example, the most powerful model for planning and the fastest and cheapest model for routine jobs. This allows flexible and efficient development without being tied to rigid processes.

Team dynamics, organization and the new roles

With more automation and AI in the team, organizational issues also arise. Teams today need to become smaller, cross-functional and more flexible in order to keep up with the speed and parallelism that AI enables. Conflicts caused by conflicting changes in the code, for example, can already be resolved well by AIs - but overview, focus and control remain human tasks. Traditional divisions such as front-end and back-end teams quickly reach their limits when everything grows closer together or an agent takes over some of the teamwork.

Changing quality assurance: from testing to feedback loops

Traditional software tests have always been about clear starting points and expected results. With AI-generated or AI-optimized code, this approach is changing. There are more variants, more feedback loops, smaller iterations. Testing is more frequent, more automated and more dynamic. Good teams move tests far forward into the development stage (“shift left”) in order to keep up with the pace of AI. If you miss this, you are left with mountains of pull requests and lengthy QA approvals.

The role of the traditional developer is shifting. In future, AI will take over many “manual operations”. The decisive skills will be to create good environments and exemplify principles that promote quality and efficiency. This is and will remain architecture work, just one level higher - and often in conjunction with intelligent agents. Companies that adapt their teams, principles and processes now will benefit twice over from progress. The others risk slipping into chaos more quickly.