Most AI discussions confuse Large Language Models (LLMs) with AI as a whole. In this episode Eberhard sits down with Nikita Golovko — industrial AI architect and practitioner at Siemens — to challenge that view and explores what happens when probabilistic systems meet deterministic environments like industrial automation.

We clarify the differences between LLMs, Generative AI, and other AI methods — and where each actually creates value. From there, the focus shifts to architecture: how to safely integrate probabilistic AI into systems that require reliability, precision, and control.

Finally, we generalize these lessons beyond industry, showing how choosing the right tool for the right problem leads to better outcomes in any domain.

Nikita will talk at the TechRiders Summit in Cologne. Register for FREE with the rebate code ARCH-TECHRIDER-2026 .

Links

• ArcAI42 - Nikita’s AI extension for arc42

• KI = Bullshit?

• Michael Stal: Analyse großer Softwaresysteme mit LLMs

“Patterns of Anarchy” is a collection of writings published in 1966. Andrew came to it because a) Christopher Alexander quotes from it in “A Pattern Language” and b) because as a consultant and developer they are interested in different patterns of organizing.

What interested them most about this book was the section “Constructive Anarchism: Alternative Communities and Programs”. This covers the how of anarchist organization.

They will share some of the most interesting insights from their perspective as a student of socio-technical organization design. Sometimes they’ll add some commentary. And of course the question is how it relates to software engineering.

Links

• Andrew Harmel-Law “Scaling the Practice of Architecture, Conversationally” at MartinFowler.com

• Book Andrew Harmel-Law “Facilitating Software Architecture”

• Andrew’s slides from Agile meet Architecture

• Book Reed Hastings, Erin Meyer: No Rules Rules

• SE Radio 253: Fred George on Developer Anarchy

Bei der Verwendung von LLMs für Software-Architektur geht es meistens um das Erstellen von Architektur. Prof. Dr. Michael Stal ist Principal Key Expert Engineer bei der Siemens AG. In diesem Stream sprechen wir über seine Erfahrungen bei der Benutzung von LLMs für die Code-Analyse. Die Analyse großer Softwarearchitekturen und Codebasen durch LLMs scheitert unter anderem an dem beschränkten Kontextfenster der Foundation-Modelle. Wie sich diese Problematik zumindest teilweise umgehen lässt, möchte der Sprecher anhand geeigneter Techniken zeigen.

Den vollmundigen Versprechen mancher Anbieter und Influencer für generative KI steht die praktische Realität gegenüber, dass die Beschränkungen von LLMs zu vielen Problemen führen, ganz besonders beim Softwareengineering. Die Analyse großer Softwarearchitekturen und Codebasen durch LLMs scheitert unter anderem an dem beschränkten Kontextfenster der Foundation-Modelle. Das gilt im Umkehrschluss auch für deren Generierung. Wie sich diese Problematik zumindest teilweise umgehen lässt, möchte der Sprecher anhand geeigneter Techniken zeigen.

Links

• Michael Stal bei LinkedIn

Der fachliche Schnitt eines Systems entscheidet darüber, ob es langfristig änderbar bleibt. Doch wie findet man einen sinnvollen Schnitt, ohne sich direkt in die Komplexität von Domain-Driven Design zu stürzen?

In dieser Episode schauen wir uns die Independent Service Heuristics (ISH) aus dem Team-Topologies-Umfeld an. Sie liefern einfache, aber wirkungsvolle Fragen, um zu beurteilen, ob ein „Ding“ als eigenständiger Service funktionieren kann.

Wir diskutieren, wie diese Heuristiken helfen, Domänengrenzen greifbarer zu machen, warum sie besonders gut mit Business-Expert:innen funktionieren und wo ihre Grenzen liegen. Ein pragmatischer Ansatz für alle, die bessere Services schneiden wollen – ohne sich in Abstraktionen zu verlieren.

Links

• Independent Service Heuristics auf der Team Topologies Webiste

• Independent Service Heuristics Github Repo

• Wir bauen eine Software-Architektur - Struktur der Lösung

• Nick Tune about Architecture Modernization

• Nick Tune - Legacy Architecture Modernisation With Strategic Domain-Driven Design

This episode was streamed live from Agile meets Architecture conference.

In this episode, we discuss the multi-year journey of Circle K’s eMobility organization as it scales to support growth from Norway to European and global markets.

The eMobility organization began as a small team focused on validating the electric vehicle (EV) charging business in Norway. However, due to its success, it quickly had to shift from “validating to scaling” and expand to various countries and multiple products in an industry that is still in development.

Throughout the episode, Eduardo and Guro will share valuable “mistakes”, lessons learned, experiments, methods, and practices we have employed during this journey. We will particularly emphasize the importance of breaking down functional silos within the organization as a means to support sustainable scaling. Initially, we focused on overcoming the Product and Technology silos. Still, in time, we went further to develop truly cross-functional value streams, also involving and continuously engaging with marketing, sales, operations, and other disciplines, with the goal of defining the best ways to support the activities necessary for rapid and sustainable business growth.

Eduardo and Guro have employed various ideas and techniques, including Domain-driven Design, Team Topologies, Wardley Mapping, and others. However, you will see that there are no silver bullets. The secret is embracing this as a continuous improvement process, involving people with knowledge and expertise, maximizing learning, and empowering value streams and their teams to drive the necessary design and decision-making with a clear long-term vision.

• Architecture Modernization Enabling Team

• Independent Service Heuristics (ISH)

  • https://github.com/TeamTopologies/Independent-Service-Heuristics

  • https://teamtopologies.com/news-blogs-newsletters/2024/8/7/newsletter-ish-enhancing-modularity-and-autonomy

• Core Domain Charts

• Susanne Kaiser: Architecture for Flow

This episode was streamed live from Agile meets Architecture conference.

We all know it - our team has become too big, meetings take too long, half of the conversations don’t apply to our work, and the sprint goal is now “finish all stories in the sprint”! The classic textbook and the chatbot are certain: The team should be split!

And this is indeed the optimal solution. But real life isn’t a textbook, and our resources aren’t infinite. What if instead of slicing to be a-two-pizza-team, we asked the question: “What do we actually need to work well together?”

After over 4 years working with several large data science and engineering teams that wrestled with multiple variations of the same problem, we’ve resisted the urge to split by the book.

Instead of insisting on the one right way, we want to show you how tuning in, listening, and deliberately choosing the solution, can bring back the fun, ease and coveted efficiency we all are after.

That could mean: changing who does what in the team, redrawing team boundaries, or combining pragmatic approaches of multiple organizational design systems like LeSS, Team Topologies, and Fluid Teams.

The trick is to stop chasing the perfect model and start designing something that actually fits both the team’s culture and unique problem domain. Think of it like tailoring a suit: it has to fit the people wearing it, not just look good on a cover.

Der Informatik-Pionier Peter Naur formulierte 1985 in seinem Aufsatz “Programming as Theory Building” die These, dass Programmieren im Kern bedeutet, eine Theorie zu entwickeln – ein tiefes Verständnis eines Problems und seiner Lösung.

Diese Perspektive erklärt, warum Änderungen an bestehenden Systemen so schwierig sind, wie Legacy-Software entsteht und weshalb iterative Softwareentwicklung so wirkungsvoll sein kann.

In dieser Episode diskutiert Eberhard Naurs Überlegungen und setzt sie in Beziehung zu aktuellen Herausforderungen der Softwareentwicklung – etwa zur verbreiteten Vorstellung im Kontext generativer KI, Programmieren bestehe primär lediglich im Erzeugen von Code.

Links

• Programming as Theory Building

• Prof. Christiane Floyd zu “menschenzentrierter Software-Entwicklung”

• KI = Bullshit

• Software-Entwicklung = Lernen?

In dieser Episode spricht Lucas Dohmen mit Eberhard Wolff darüber, wie man Anwendungen aus dem Cloud-Angebot großer Hyperscalers wegmigriert. Er berichtet dabei aus der Praxis: Gemeinsam mit dem Team von fejo.dk, einem der meistgenutzten Portale für Ferienhäuser in Dänemark, hat er die Anwendung von Amazon Web Services (AWS) in die Hetzner Cloud umgezogen. Lucas erläutert, wie sie dabei vorgegangen sind, welche Vorteile es gibt, welche Herausforderungen sie lösen mussten und wie ein solcher Weg typischerweise aussieht.

Links

• Hyperscaler-Exit bei SWAGLab

• Frage zu Hetzner bei Mastodon

• Frage zu lokalen Points of Presence bei Mastodon

• Serverless Architektur mit Sascha Möllering

• LinkedIn Frage zu Kamal