Carl-Christian Sautter

Also: Ticket.

Tage später die Antwort. Kein Fix – sondern eine Erklärung, was du besser machen könntest und wo die Fehler vermutlich liegen. Du prüfst alles. Doppelt. Alles okay. Ping zurück.

Ab diesem Moment wird nur noch Ping-Pong gespielt.

Niemand ist mehr auf Fehlersuche. Zurückgespielt wird nur noch die Verantwortung.

"So etwas gibt es doch heute nicht mehr."

Doch. Öfter, als man denkt.

Was mich zu der Frage bringt:

Gibt es sie noch, die Engineers, die den vollen Stack beherrschen? Die einen Fehler ganzheitlich betrachten – über Team- und Systemgrenzen hinweg?

Oder haben wir uns so tief spezialisiert, dass am Ende niemand mehr das Ganze sieht?

Wie erlebt ihr das?

#SoftwareEngineering #PlatformEngineering #DevOps #Fullstack #Engineering

So I turned it into a public demo project for learning: a dev-container where an AI agent runs black-box pentesting, white-box code review, and an orchestrator that merges both into one prioritized report — driving real tools like OWASP ZAP, nuclei, sqlmap, and semgrep.

The results were surprisingly good — the whole workflow was powered by Claude Opus 4.8. I even validated it against public endpoints — with the owners' written consent, fully legally. ⚖️

Still somewhat tailored to the OWASP Juice Shop case, but built to be adapted.

This project is a hands-on showcase of agentic coding and security workflow design — turning an LLM from a code assistant into an autonomous agent that orchestrates real tooling end-to-end.

Repo's public 👇 What would you have the agent test next?

https://lnkd.in/eDkzZmUX

Which feels slightly weird, because the last time I really built websites hands-on, “agentic” mostly meant asking a colleague if they had already fixed the CSS.

Years ago, I started my software engineering career by building websites.
TYPO3 was one of the first CMS systems I learned back then. And setting it up was absolutely manageable, of course… but let’s just say it was not exactly a “one prompt and coffee” situation.

Over the last years, I haven’t worked directly in classic website building that much.
And now I’m coming back to it in a very different way:
Claude Code is building the website.
The content is generated and inserted by an agent.
And even the TYPO3 setup is handled by a TYPO3 agent.

So I went from:
“Let me configure this manually.”
to:
“Let me supervise a small group of agents while they negotiate with TYPO3.”

Crazy times.
And honestly, a pretty fascinating full-circle moment.

Denn so sehr ich Bahnfahren eigentlich gerne als Arbeitszeit nutze: Sobald Videocalls ins Spiel kommen, wird es schnell zum Resilienztraining.

Ich bin immer wieder überrascht, wie häufig es zu kurzen Aussetzern kommt und wie wenig geeignet das Internet im ICE für stabile Calls ist.

Diesmal habe ich auf meiner Bahnreise Speedify getestet — mit zwei Smartphones als Backup-Leitung:
📱 Vodafone
📱 Telekom
Mein Eindruck: Das Setup reduziert die vielen kurzen Ausfälle spürbar, auch wenn es sie nicht komplett beseitigt. Ein Funkloch bleibt nun mal ein Funkloch.

Aber durch Redundanz und sehr schnelles Failover kommt deutlich mehr Stabilität in die Verbindung. Man nähert sich zumindest etwas, das man vorsichtig als „nutzbare Verfügbarkeit“ bezeichnen könnte.

Natürlich hängt das stark von Strecke, Auslastung und Netzabdeckung ab. Aber da ich diese Probleme inzwischen auf fast jeder längeren Strecke irgendwann erlebe, bleibt die wackelige Internetanbindung im ICE ziemlich nervig, wenn man versucht, ernsthaft zu arbeiten.

Oder anders gesagt:
Remote Work funktioniert im Zug erstaunlich gut — bis das Internet merkt, dass es auch mitfahren muss.

I burned through all my included LLM tokens. 🔥

To be fair, I have three different plans for different use cases:
One for work.
One provided by a client.
One private.

So yes, I still had options.

But it forced me to think more seriously about usage, cost, and fallback strategies.
Because what happens near the end of the month when you still need to work, but one of your LLM providers has reduced token budgets, changed pricing, or simply says:
“See you next billing cycle.” 😅

Sure, you can buy additional token packages.
But honestly, many of those options feel expensive pretty quickly.

So I gave OpenRouter a try.
For those who do not know it: it is basically an intermediary platform that gives access to many relevant models through one interface.

What I liked:
You can test different models easily.
You only pay for actual usage.
There is no subscription lock-in.
And the dashboard gives you near real-time visibility into usage and cost.
Also, not every task needs GPT-5.5 or Claude Fable 5.

For basic coding support, summarization, refactoring, documentation, or “please clean up this messy thing I just wrote” tasks, cheaper models can already be surprisingly useful.
I tried models like DeepSeek V4 Flash and GLM 5.2, and for many basic tasks they were more than good enough.

OpenRouter might become my new go-to place when I want to try a specific model that is not available in Codex or Copilot.
And honestly: it is a pretty nice fallback when the monthly token budget is already gone.

Curious how others handle this:
Do you optimize your LLM usage already, or are you still in the “tokens are infinite” phase?

I mean… efficient.

Mine:
run skill "create commit"

It checks the diff, proposes a commit message, and waits for approval.
Tiny automation, but exactly the kind that removes one more repetitive micro-task from the day.

Honestly, it may be harder than naming my daughter.

With a child’s name, you think about meaning, sound, family opinions, and whether it still feels right years later.

With a software project name, you add a few extra constraints:
- Is it searchable?
- Is the domain available?
- Is it unused on GitHub?
- Is it already taken on package registries?
- Does it sound like a serious tool?
- Is it memorable without being gimmicky?
- Does it still make sense when the project grows?
- Will people know how to spell it after hearing it once?

And then, when you finally find something that feels right, you hesitate to share it too early because you don’t want to dilute the brand before it even exists.

Naming really is one of the hardest parts of building.

For people who have launched open-source tools, products, or side projects:
How did you know when the name was “good enough”?

Yesterday’s thunderstorm caused a short power outage — just a few minutes, but enough to remind me how incredibly stable Germany’s power grid usually is.
Honestly, I can barely remember the last outage before that.

And maybe that reliability is exactly why backup power is so easy to ignore.

Over the last 1–2 years, my homelab has been steadily growing again.
Ironically, about 5 years ago I moved most of my workloads into the cloud because it was simply easier to maintain and gave me a better cost/value ratio.

But over time, I started bringing more systems back under my own control — especially for learning, automation, AI workloads, and self-hosting projects.

At the same time, I’m constantly starting new projects that require different kinds of hardware resources.
And for experimentation, learning, AI workloads, automation, or self-hosting, having a flexible and “growing” homelab is often just easier.

For experimental systems, I honestly don’t care too much if something crashes unexpectedly.
But some systems slowly became the important ones:
storage, backups, automation, and services containing data I actually care about.

So when the power went out yesterday, it reminded me how easy it is to forget the layer underneath everything else: reliable power delivery.

That’s how I finally invested in one of the most unsexy homelab components possible:

A UPS (USV).

The thunderstorm reminded me that the boring infrastructure parts are often the ones quietly protecting everything else.