Blog

Trust but Sandbox

Coding agents can now do (almost) everything. They read and write every file you have access to, they run arbitrary commands, talk over the network, read your environment variables, and drive your browser with all your cookies. In the spirit of an agentic workflow, where we let agents work as autonomously as possible, that is actually the point, because it is exactly what makes them useful. Even so, we should find a way to gain more control over what happens in there.

Beyond Claude & GPT

Sure, we enjoy trying out the latest frontier model as much as anyone. But in day-to-day project work, what matters most is control over cost, data, and availability. That’s why we increasingly rely on open-weights models paired with open source tooling. Here’s what our setup looks like.

Alternatives to U.S. Artificial Intelligence

Why rising AI costs and digital sovereignty concerns are pushing me to look beyond Claude and GPT

Faster Is Not Better: On Experimentation Culture and What It Costs

Faster experimentation sounds like an obvious advantage. But experimentation culture is not a universal paradigm. It is a product of a very specific context. And the people who bear the heaviest cost of poor product development never appear in any experiment.

The Agentic Trio

The Product Trio separated discovery from delivery because delivery required a larger team. Agentic development removes this constraint. That changes what a product team can look like.

Using the space, not optimizing the treadmill

Why agentic development needs product discovery

Nebu: Self-made sovereignty

There are plenty of open-source alternatives to Slack. None of them deliver real sovereignty. Community editions are deliberately pared down so no company can run them in production, while enterprise features sit behind commercial licenses. This article shows why “building it yourself” is no longer a utopian idea today, but a realistic option for anyone with a product vision and a bit of DIY courage.

From Vibe Coder to Agentic Engineer

I no longer read every line of code that’s produced in my projects. I think most people who work intensively with AI agents feel the same way. When an agent produces a thousand lines in ten minutes, line-by-line review is no longer realistic. At the same time, that code ends up in production, and someone has to take responsibility for it.

Hail Mary: Why domain knowledge cannot be extracted from experts

Developing with AI through the cognitive Lens

The Right Kind of Hard

More structure, more output, more exhaustion. Spec-driven development has helped me get better results from AI agents – and drained my energy in the process. About dead weight, false clarity, and hidden costs I didn’t plan for.

Where AI Helps (and Hurts) Across Different Coding Scenarios

Where does LLM-assisted software development improve developer productivity—and where does it fall short? Instead of treating AI in software development as a one-dimensional productivity booster, we explore the question across several dimensions suggested by a Stanford-adjacent study: project maturity, task complexity, and the popularity of the programming language. The goal is to create a more realistic expectation baseline for both software developers and leaders—well away from today’s hype.

Spec-Driven Architecture: When Agents Build, Architecture Must Speak

SDD describes how a system is built. SDA describes how systems fit together

REST Beats MCP

Instead of using dedicated APIs, agents can operate existing web applications directly. Like humans, they rely on the most consistent implementation of hypermedia and benefit from existing context, validation, and access controls. Using an expense report as an example, I show how agents can automate complex, context-dependent tasks—without having to implement new APIs.

Spec-Driven Development is Domain-Driven Design’s Impatient Cousin

Spec-driven development tools like BMAD promise to fix the requirements bottleneck that agentic development teams face. But is it actually a good fit for your organisation?

Agents good in the end?

A story of unleashed productivity and mental load

Understanding AI Coding Patterns Through Cognitive Load Theory

AI coding assistants either severely harm learning or enhance it beyond manual coding—depending entirely on interaction patterns. A recent study reveals six distinct patterns with dramatically different outcomes. We’ll examine why through cognitive psychology and explore the study’s limitations.

Ein Überblick der AI-assistierten Softwareentwicklung in 2026

Birgitta Boeckeler (ThoughtWorks), Heinrich Hartmann (Zalando) und Sven Johann haben sich über den aktuellen Status der AI-assisted Software Entwicklung unterhalten: was ist wirklich Wichtiges in 2025 passiert? Welche stabilen Muster können wir ableiten? Was gibt Orientierung und Einordnung? Und natürlich ein paar Einblicke in den Future of Software Engineering Retreat in Utah an dem Birgitta teilgenommen hat. Die Unterhaltung gibt es natürlich auch als Podcast. Dieser Artikel ist aber eine menschlich kuratierte Zusammenfassung.

Neuland reloaded

I sandboxed my coding agents. Now I control their network.

I want my AI coding agents to work independently, but I don’t want them to have unrestricted access to the internet. In this post, I describe how I routed all network traffic from my development sandbox through a strict proxy allowlist, allowing only a small set of explicitly approved domains. This setup finally gave me enough confidence to loosen the guardrails without constantly staying in the loop.

From Vibe Coder to Code Owner

AI agents generate thousands of lines of code in no time. If you want to fully leverage their potential, you can no longer review every single line – but you’re still responsible for the software. How do you take ownership of code you haven’t fully read? The answer lies in the agent harness: a system of deterministic checks, AI reviews, and targeted human review that enforces quality instead of hoping for it.

Let’s Not Normalize Insecure AI Assistants

How the Lethal Trifecta Makes Today’s AI Assistants Unsafe by Design

What to Do While the AI Is Thinking

AI can write code while we wait — but that doesn’t mean we suddenly have free time. Whether we should think, multitask, review, or rest depends on what kind of work we are doing.

AI and Elaboration: Which Coding Patterns Build Understanding?

AI tools let you complete coding tasks without connecting new information to your existing mental models—a cognitive process known as elaboration that is crucial for building understanding. But some AI interaction patterns preserve this elaboration while others bypass it entirely. Let’s explore what elaboration is, why it helps with learning, and how we can use AI tools in a way that helps with this process rather than circumventing it.

Fetch-Tools vs. Browser-Rendering in Agenten-Setups

I sandboxed my coding agents. You should too.

LLM coding agents are extremely powerful because they can run programs on our computers using our permissions. However, this same power also makes us very vulnerable. It only takes one mistake or one prompt injection to compromise the whole system.