I Run a Travel Magazine. My Editors Are Agents.

In 2008 I built a travel website called Hamba, in PHP, the way we built everything back then. It lived, it aged, it was forgotten.

Years later I was cleaning up old disks and found the codebase. For fun, I had Copilot port it to .NET, and here's the strange part: it worked. The whole site, alive again in an afternoon. But looking at it, I realized something more interesting: the code was fine, the thinking was outdated. It was a website built for a world where humans wrote every page by hand. So instead of shipping the resurrected version, I asked a better question: if I built Hamba today, from zero, in the age of agents, what would it be?

The answer is hamba.nl, a Dutch travel magazine where the editorial pipeline is run by a team of AI agents. It's a real production system, not a slide demo. And running it for a while has taught me more about working with agents than any tutorial, so let me give you the tour.

The newsroom

Think of it as a small newsroom where the staff happens to be digital, and every employee has exactly one job.

There are agents that do the thinking: research, planning what a piece needs, deciding structure. There are writer agents, one for Dutch and one for English, and I'll be honest: the prompt engineering for those two took more work than anything else in the system. Getting an agent to write is easy. Getting it to write like your magazine, in two languages, without slipping into the generic AI voice, is a craft. There's an agent that makes the images. And there's a validator that checks every piece against the editorial rules before anything moves forward, because an editorial pipeline without a quality gate is just a content hose.

At the end of the pipeline sits the most important part: a human. Nothing publishes without approval. More on that in a minute, because it's not the disclaimer it sounds like. It's the design.

Hire the model for the role

Here's something running this taught me that demos never show: you don't pick one model, you hire a team.

The model that's great at thinking through what an article needs is not the model I want writing the prose. The one that writes beautiful Dutch is not automatically the one for images. So each role in the newsroom runs the model that's best at that job, and I test candidates the way you'd interview new hires: give them the actual work, compare the results, pick per role. Which models? That answer changes every few months, and that's exactly the point.

Build the newsroom so you can swap an employee without rebuilding the office.

If that sounds like my hiring advice for humans, that's not an accident. One job per agent, the right hire per role, and judgment stays with whoever has actually done the work.

What agents are like as colleagues

After months of running this team, here's my honest performance review.

Great at: volume, consistency, and never having a bad Tuesday. The pipeline produces at a pace no human newsroom this size could match, in two languages, every piece following the same editorial rules.

Terrible at: taste. An agent doesn't know that a piece is technically fine but boring. It doesn't know that this opening has been used a hundred times, or that a place description sounds like it was written by someone who has never been anywhere. That judgment, the difference between correct and good, still lives entirely on the human side of the pipeline.

That's why the approval step is the design and not a disclaimer. The agents take the volume. The human keeps the taste. Sound familiar? It's the same line I draw everywhere: automate the task, never the point. The point of a magazine is that it's worth reading, and worth reading is a human judgment.

You can't manage a team you can't see

The least glamorous lesson, and maybe the most important one for anyone building multi-agent systems: observability is not optional.

When a piece comes out wrong, I need to know which agent did what, in what order, with what input. So every run is traced, and I built a visualizer that can replay a full pipeline run step by step. It's the agent version of an editorial meeting: what happened, where did it go sideways, who needs better instructions.

I've written before that trust is protected in spreadsheets. In an agent system, trust is protected in traces.

If you can't see what your agents did, you don't have a team. You have a mystery that occasionally publishes.

Why bother?

A fair question: why run a real magazine instead of just building demos?

Because production teaches what demos can't. A demo has to work once, on stage, for ten minutes. A magazine has to work every day, with real content, real edge cases, and real consequences when the Dutch writer agent suddenly develops a strange new habit. Every hard lesson in this post came from the gap between those two. I've said before that you should build the thing before you demo the thing. hamba is me doing that with the biggest thing I talk about on stage: a workforce of agents.

And the meta-lesson, the one that keeps coming back in everything I write: the more the agents produce, the more valuable the human at the end becomes. I didn't automate myself out of a magazine. I automated myself into the only job that matters there: deciding what's good.

The 2008 version of Hamba needed me for everything, so it died when I got busy. The 2026 version needs me for one thing, so it lives.