AI agents are all the rage. But not because they're particularly useful. Rather, because many want to hide their lack or half-knowledge behind buzzwords, or simply don't know any better. Software development without AI is nonsense, but with agents it's a dead end.
Introduction
Agents have become a buzzword for some time now. An AI agent is an autonomous program that uses AI models to accomplish complex tasks.
For example, an agent can perform research tasks or remote control programs, thereby relieving humans.
For an AI agent to be able to perform its task, it requires broad access rights. As always, great opportunities also come with great risks.
In the context of programming applications, websites, or apps, a AI-agent is like a new employee who can code. You give the agent an instruction on what to program. Then, the agent starts working and independently programs what you commissioned.
The agent creates new files to store the program code, or it modifies existing program code. Depending on the mode, the agent asks for your consent every time the source code is changed. After a few seconds or minutes, depending on the task, the desired program is finished.
Advantages and Disadvantages of Agentic Coding
Benefits
The advantages of AI programming using agents are obvious. What previously took hours, days, weeks, or sometimes even a year can now be accomplished within seconds or minutes.
Not only does the time for completing programs drop rapidly, but it's also crucial that the human energy expenditure is significantly reduced. It's simply much easier to perform a very simple task than one that requires strong intellectual demands. Many fail even because they don't start with challenging tasks at all. With AI-programming, this hurdle completely disappears.
In late 2025, a developer at a Big Tech company managed to achieve a result within an hour using AI programming, a result that had previously taken several highly paid developers working together for a full year.
The main advantage of agent-based AI programming is primarily for novice programmers or developers with moderate abilities. What was previously impossible is now possible.
A participant in a Dr. GDPR workshop on AI programming told me a few weeks later on the phone that he had programmed himself a MP3 player of his choice. He had never programmed anything noteworthy before. He also managed to develop a program for the Raspberry Pi that analyzes his heating system, even though he had never programmed anything on the Raspberry Pi before.
Drawbacks
Agentic Coding, however, also has some drawbacks. These primarily affect professional software development.
Agentic coding is primarily suitable for toy projects.
For professional software projects, agents should not be used.
Agents take complete control of what happens. In agent-based programming, the agent autonomously creates source code, modifies existing program code, and reports “Done” at the end. An experienced developer might still be able to follow what happened with very careful scrutiny – until the program becomes too extensive or time is too short.
The complete loss of control means on one hand that humans no longer understand the software programmed by the agent. Security gaps arise and remain undetected, and dangerous accesses can bring entire system landscapes to a standstill.
If an individual uses an agent to program an app, then the process is, at best, as follows:
- The DAU (Dumbest Assumable User) gives the AI agent a task, such as: “Write me an MP3 player that runs in my browser and lets me manage a playlist.”
- The AI agent creates the MP3 player in about 2 minutes.
- The user starts the MP3 player. Now there are several possibilities:
- The MP3 player is working, but the user has desires for modifications (luxury problems).
- The MP3 player is functioning, but it does not provide the user with (serious) functionality.
- The MP3 player is not working or has serious, show-stopping errors.
What is guaranteed not to happen: The AI agent does exactly what it is supposed to.
As is known from practice, a technically unskilled user reaches a point where they can no longer proceed. The user simply does not understand what the agent has produced. The user can only repeatedly give new instructions to the AI agent. The conversation becomes longer and longer. The agent's memory (context) becomes fuller and fuller. The results get worse instead of better.
Software development with AI agents is thus a one-way street with a dead end.
AI agents crash systems full force into a wall and eventually destroy them.
Except, if humans intervene.
In addition, there are significant risks associated with using AI agents. For example, the agent can create a program that causes damage (delete files, cripple networks, exfiltrate passwords, cause system crashes, etc.). Programs that work incorrectly are less easily detected. In a financial accounting AI, it would be better if all numbers were correct. 99% is not enough here. Everyone knows what it means to have a 9 instead of a 1 before the comma, or to move the comma one place to the right or left.
The Solution
What to do if you don't use agents for programming? Quite simply:
AI programming yes, agents no.
What does this mean concretely? The core features of a good solution for software development are:
- Absolutely (also) use AI systems to develop software!
- By no means use agents!
- Involve the developer.
- The developer is responsible. They alone!
- Establish a process that effectively utilizes AI to develop highly efficient programs
My name is Klaus Meffert. I have a doctorate in computer science and have been working professionally and practically with information technology for over 30 years. I also work as an expert in IT & data protection. I achieve my results by looking at technology and law. This seems absolutely essential to me when it comes to digital data protection. My company, IT Logic GmbH, also offers consulting and development of optimized and secure AI solutions.
