The Vision of Agentic Software Development

Recently I got a sneak preview about the future of software development in an interesting show – the world according to tool vendors …

The new mantra is all around orchestrated agents, working 24×7, conducted by a single human developer, with access to all knowledge being used to train the AI model, fixing security flaws faster than ever and generate new applications quicker and more reliable than ever.

What Are We Actually Producing?

But what exactly do they produce in the first plave? Well, Terrabytes of code. But relevant and desperately awaited by human mankind? This is an unanswered question for now, as it depends on the context, all of the functional and non-functional requirements handed to the skill ful hands of agents.

Innovation, Speed and Energy Consumption

Don’t get me wrong, I’m excited about the progress that we saw across the past decades in the area of software development. The current speed of innovation in the AI area is breathtaking, and it still seems to accelerate every day, overwhelming most of us to keep up to date!

Speed, however, is not a guarantee to get a unique value back – lots of results are duplicate results and not innovation. The current hype around AI generates a record consumption of tokens In the end we are not burning tokens, but rather energy in data centers that force some of the power suppliers erect new power plants near the new AI data centers. Humans love to play, and it will take a while before we really identify the use cases with the highest return.

Hallucinations: A Persistent Challenge

There is a different aspect going hand in hand with the rising consumption of AI.Although mightier than ever, they still hallucinate here and there. A human being would be locked up in a psychiatric centre. As an AI agent, you’re allowed to continue with your work and will not be taken into custody. The percentage of hallucinations is going down, relative to the amount of code produced. But there are critical systems and application with zero tolerance to mistakes that might impact live and healthy of users, so you better not miss a single „vision“. The more we rely on AI without keeping the skills to decide what is good and bad code, we either have to move back, or we have to live with it.

The Business Model Behind the Hype

But why do vendors propagate orchestrated agentic AI so heavily? Because it offers a different business model. Instead of subscribe for seats of your human developers it will be the consumption that matters most. Seats are limited by the number of people, consumption can be endless. The message is: „the more tokens you us, the higher the value you get“ What is left out is the fact that the bill is getting higher as well. No problem, if you really get value back that your enterprise can monetize, but this is something that agents cannot tell you (yet). It is still about humans to decide on the context and what agents are supposed to do.

Asking AI About the Risks of AI

To me it has always been a matter of honour to write my blogs wihtou AI support. This time I had to make an exception, as I was asking my Gemini „Why should humans not use AI agents?“.

Gemini’s Answers

The answers I got (original screenshots can be provided upon request):

Data Protection RViolations – AI agents need deep access to personal data and passwords

Loss of control – autonomous systems can cause unpredictable mistakes

Security flaws – hacker can manipulate agents with prompt injections

Job losses – full automation in certain areas

Social isolation – loss of social skills in human interaction

Loss of skills – critical analysis and the competence to solve problems can shrink

A Question of Balance

It is the first time that a technology is warning to be used excessively. I think we need to think twice and have our skills and capabilities trained. Let them do what they are good at, routine work that tires humans, but leave the creative and responsibility to the people.

Author: Rainer Heinold