Jakob and ADOPS.online – AI-Native Enterprise Marketing Automation: AI Coding interview

The question occupying me most right now is this: How do we enable intelligent systems to make decisions independently without losing safety, data quality, or control?

One focus is our Evidence-Backed Network Identity Fabric, or EBNIF for short. It is meant to reliably identify and assign companies based on verifiable data and network signals. That may sound unspectacular at first, but technically it is highly complex – especially when different sources supply contradictory information.

In parallel we are developing a Container Twin Set architecture. The idea behind it: prepare changes in an isolated environment, fully verify them, and only then roll them out in a controlled way into live operations.

AI Coding enables me to work on such demanding architecture topics at a speed that used to require an entire development team.

But that is exactly where the challenge also lies.

An AI can develop complex functions in a few minutes. Whether those functions work reliably with the rest of the system is an entirely different question.

This experience now shapes every architecture decision I make.

My day-to-day development work has fundamentally changed. I now spend significantly more time orchestrating tasks, making architecture decisions, and reviewing results than writing code myself.

ChatGPT is one of my most important engineering partners in that process. I use GPT models especially for architecture questions, complex analyses, and the systematic investigation of technical problems. Codex, Claude Code, and Cursor complement the coding workflow. Gemini is particularly interesting for extensive specifications and large volumes of information. Grok, Antigravity, and Hermes Agent are also among the tools I engage with as part of my multi-agent approach.

What matters most, however, is the environment in which these agents work.

Via GitHub, Remote Desktop Commander, and MCP they can access the real development environment, analyze relationships, prepare changes, and run tests.

On top of that comes an extensive system of project-specific skills. They define how agents must handle architecture, security, tests, and deployments.

Our technical foundation consists of Python, FastAPI, Next.js, React, TypeScript, PostgreSQL, Redis, Docker, and GitHub Actions.

One of the biggest challenges is state management. When multiple agents or workers want to modify the same resources at the same time, race conditions or duplicate actions quickly arise.

That is why idempotency, controlled transactions, and clear responsibilities have become indispensable for me.

I do not simply want to deploy as many AI agents as possible. I want them to work together reliably.

The numbers now show quite impressively the scale the project has reached.

Since early February 2026 I have been working on ADOPS.online every day. By my estimate that now amounts to around 2,700 development hours. My GitHub profile alone records more than 8,300 contributions in 2026.

AI consumption is also substantial: In a single OpenAI API project, around 34.7 million tokens were processed in October alone across more than 31,000 model requests.

On top of that come further models, coding tools, and development environments. The complete cross-provider token balance has not yet been finalized.

The GitHub and OpenAI invoices that can be traced so far are in the four-digit US dollar range. That figure also includes CI/CD and development services, however. Additional tool and infrastructure costs are still being consolidated.

My workflow today is strongly agent-based: specify requirements, distribute tasks, implement changes, run tests, and review results.

That sounds straightforward. Reality was often far more chaotic.

One of my most painful experiences: An AI agent successfully fixes a bug – and in doing so damages something that previously worked perfectly.

That is exactly why I have increasingly aligned my development process around isolated changes, regression tests, and controlled releases.

Today I am less interested in how many tokens I consume. I am interested in how much truly working software comes out of it – and how reliably it still works tomorrow.

I have been involved in software development for more than 25 years. I started in digital marketing, with tracking systems, ad servers, and technologies that later evolved into ever more complex data and automation architectures.

With GPT-3 and early Codex tools I ran my first coding experiments in 2021/2022. At the time it was mainly about individual scripts, functions, and smaller automations.

In 2024 that changed fundamentally.

Suddenly I could discuss entire software architectures with AI, develop complex modules, and analyze problems that went far beyond simple code completion.

For me that was a real turning point.

At the same time I realized how valuable my prior experience is in that context.

Because AI can program impressively fast. It can just as quickly implement a poor architecture if nobody asks the right questions.

After 25 years I know one thing above all: The most dangerous software bugs are often the ones you do not even notice at first glance.

My focus has therefore shifted. Today I think more in terms of systems, dependencies, risks, and technical contracts.

AI takes on more and more of the implementation. Responsibility for the result remains with me.

And it is precisely this combination of experience and new technological capability that I find incredibly fascinating.

Yes, I am convinced of that. But hardware confronts AI agents with challenges that we barely know in this form from software development.

Even today, AI methods support chip design, circuit drafts, PCB layouts, and complex simulations.

For me the big difference lies in the development cycle.

I can roll software back. Not physics.

If I discover a faulty build at ADOPS.online, I can discard it, correct it, and test again. With a chip or a physical assembly, weeks or months can pass between design and real test results.

On top of that come thermal loads, electromagnetic effects, material tolerances, and manufacturing deviations. A design can look excellent in simulation and still fail in reality.

That is why I do not see the greatest challenge in whether AI can design hardware at all.

The real hurdle is securing the entire path from digital design to a reliably functioning physical product.

To do that, AI agents must work together with simulation tools, formal verification methods, measurement systems, and manufacturing processes.

And that same idea occupies me at ADOPS.online as well.

Intelligent individual agents are impressive. It only becomes truly revolutionary, however, when we can connect them into reliable systems that work with one another.

I believe one of the next major technological breakthroughs will emerge here.