I update my blog with AI agents

If there's one thing I've learned running IT for years, it's this: if you do the same manual task twice a week, you should automate it by now. That's why this blog, like most of my workflows, runs on AI agents.

I'm not talking about a chatbot bolted onto a website. I mean real automation: event-driven workflows that watch for triggers, decide what to do, and act without me touching a keyboard.

The stack

Two tools carry most of the load:

• n8n: my self-hosted workflow engine. It handles the heavy orchestration: webhooks, API calls, branching logic, and hand-offs between agents.
• Power Automate: the glue for everything Microsoft-centric, since most of my clients live in M365.

On top of those sits the agent layer. Instead of hard-coding every decision, I give an LLM a narrow job, a context window with the relevant data, and a set of tools it's allowed to call. It plans, executes, and reports back, and if a step fails, the workflow catches it and retries with a different approach.

A concrete example: the publish pipeline

Here's a flow I run every time I finish a post. An n8n webhook receives the draft's frontmatter (title, tags, slug) the moment I save it. The workflow then:

1. Runs a lint agent that reads the draft and checks tone, structure, and technical accuracy against a style guide I wrote once.
2. Routes the result: pass → move on; fail → the agent rewrites the offending sections and sends me a diff for approval.
3. Generates the metadata (description, OpenGraph text, suggested tags) and updates the post frontmatter.
4. Triggers a build hook on the server, then a health-check agent fetches the live URL and confirms the post is up with the right title and status code.
5. Posts a short update to my social accounts via the same workflow.

From "draft saved" to "live and announced," nothing needs me unless something fails, in which case I get a message with the exact log excerpt.

Why this matters

Automation isn't about removing humans. It's about removing the mechanical parts so the human parts (judgment, writing, architecture) get more attention. Every flow I build is a small portfolio piece: it proves I can take a messy process, model it, and make it run itself.

Want to see more? I'll keep publishing the flows that survive contact with reality, including the ones that failed first.