AI Agent, n8n, and a Week of System Work: What I Built

TL;DR

This wasn't a product week. It was a systems week. Infrastructure, automations, pipelines: things nobody sees but everything depends on. I migrated a GitLab domain, set up n8n as an automation hub, launched a content pipeline live on Instagram, and worked through more than 50 new issues with my AI agent Clank. Here's what came out of it. And what still doesn't work.

The difference between visible and invisible work

As a solo founder, you have two kinds of weeks. The visible ones, where you ship a feature, publish a post, win a customer. And the invisible ones, where you build the foundation everything else stands on.

This was an invisible week. And somehow one of the most productive ones in a while.

If you build and run a product alone, you know: when the infrastructure is shaky, everything is shaky. A broken monitoring setup catches problems too late. A misconfigured CI/CD pipeline costs hours. A missing webhook means a critical event vanishes without anyone noticing.

That's exactly what I fixed this week.

GitLab migration: small change, clear signal

The smallest change carried the biggest meaning. I migrated my GitLab instance from gitlab.buchhaltgenie.at to gitlab.gotzendorfer.at. Technically, a DNS redirect and a few script updates. Symbolically, a clear statement: this ecosystem belongs to me, not to a single product.

48 repos, 5 groups, a monitoring stack, CI pipelines on multiple servers. All of it now under one roof. The 301 redirects work. All webhooks point to the right endpoints. Nobody notices the difference, which means it worked.

n8n as the automation hub

Since last week, n8n is running on my CI server as a central automation hub. The idea is simple: my AI agent Clank is good at understanding things, evaluating them, and deciding. n8n is good at watching triggers, calling APIs, and running workflows reliably. Together, they make a system that can do more than either one alone.

What's running now: when a new booking comes in via Cal.com, it lands automatically in my Notion CRM. When a lead inquiry comes through the website, it gets categorized and forwarded to Clank. When Clank completes a significant action, like creating a GitLab issue or finishing an analysis, he logs it to a Notion database so I keep the overview.

Three new workflows I prepared this week: a weekly scan for IT tenders in Austria (public procurement is something I'm exploring), a competitor monitor that searches for relevant market shifts automatically, and a pipeline that turns Sentry errors into GitLab issues when confidence is high enough.

That sounds abstract. Concretely: I sleep better now.

FeedFoundry: content pipeline live on Instagram

Since Monday, FeedFoundry, my content pipeline, is automatically posting to Instagram for BuchhaltGenie. The first image is live. No carousel, no reel. A simple post that proves the pipeline works.

Behind it are eight steps: idea pre-selection, context from the repo, format decision, text creation, image generation with Stability AI, quality check by a second model, adaptation, and finally publishing.

What occupies me most about this project: quality control for AI-generated content is not optional. A bad post does more damage than no post. That's why the pipeline has a quality threshold of 0.8 on a scale from 0 to 1. Anything below that doesn't get published, no matter how many iterations it takes.

EventDrop: paying down technical debt

EventDrop, my event management tool, reduced several pieces of technical debt this week. Cover images for recap videos now work. Statistics queries were migrated from an inefficient full-table scan to an optimized database function. And a long-standing CORS bug in the video upload has been documented and has a concrete solution concept.

These aren't features that show up in an App Store screenshot. But they're the difference between a system that breaks under load and one that just runs.

What AI agents can actually do today

I get asked often what it's like to work with an AI agent. The honest answer: better and more limited than most people expect.

Better, because Clank handles things in parallel that I'd forget, because he doesn't get tired, and because he's faster than me on clearly defined tasks. This week he created more than 50 GitLab issues, rolled out webhook configurations on 8 repos, structured n8n workflows as JSON, and tested n8n instances.

More limited, because he still hits system boundaries. n8n 2.9.4 has a bug in webhook registration via API. Clank found it, documented it, and built a workaround, but it can only be resolved through the UI. That's not a failure, it's reality: AI agents are force multipliers, not magicians.

What I've learned from this: the most valuable moments aren't when Clank handles something I would've handled anyway. The most valuable moments are when he handles something I would've forgotten, or never gotten around to.

What's coming next week

Three things at the top of the list. First, AIDL2026, a public IT procurement process in Austria where I want to position myself as a sub-contractor. The deadline is April, preparation starts now. Second, the goetzendorfer.at relaunch by end of March. Third, getting the n8n workflows prepared this week into active state.

Systems weeks aren't glamorous weeks. But they're the weeks that make the difference.