What I learned running a newsroom of twenty agents

I built a newspaper with no journalists. Not a demo, not a prototype in a notebook, a real one that publishes several times a day, on its own, and once earned organic press coverage in two countries without a single human in the newsroom.

Under the hood it is around twenty AI agents arranged like an actual editorial hierarchy. An editor in chief sets direction. Section editors decide what gets covered and shape the angle. Specialist agents research, write, and fact check. A publisher agent handles the mechanics of getting it live. It runs on a schedule, defends itself against prompt injection, stays inside a strict budget, and corrects itself when a reader flags a hallucination.

Building it taught me things that no tutorial about agents will tell you, because tutorials stop at the part where one agent answers one question. The interesting problems all live past that point.

The first lesson is that orchestration is the whole game. A single agent doing a single task is easy now. Twenty agents that have to hand work to each other, agree on what is true, and not step on each other is a different kind of problem. It looks less like prompting and more like designing an organization. Who decides. Who reviews. What happens when two agents disagree. I spent far more time on the flow between agents than on any individual prompt.

The second lesson is that autonomy without guardrails is just a fast way to fail. An agent left alone will confidently do the wrong thing at scale. So most of the real engineering was not the clever part, it was the boring, essential part: budget ceilings so it cannot spend itself into the ground, injection defense so a malicious page cannot hijack the pipeline, and a verification loop so a hallucination gets caught and corrected instead of published and forgotten. The magic is not that it runs on its own. The magic is that it runs on its own safely.

The third lesson is that you cannot improve what you cannot see. With one agent you can eyeball the output. With twenty, working in parallel, on a schedule, you are blind unless you build the ability to watch them. Where did the cost go. Which agent produced the weak paragraph. Why did this story come out wrong. Observability stopped being a nice to have and became the difference between a system I could actually run and a black box that occasionally embarrassed me.

The fourth lesson is the one I did not expect. Building a fully autonomous system made me appreciate human judgment more, not less. The agents are fast and tireless and genuinely good at their pieces. What they lack is the sense of when something is off in a way the rules did not anticipate. My job, in the end, was to encode as much of that judgment as I could into the structure, and to stay honest about the parts I could not.

People ask if this replaces newsrooms. That is the wrong question, and it misses what the project actually is. It is a working answer to a harder question: how do you make a large number of AI agents do real, consequential work together, over time, without a human catching every mistake. That question is going to matter far beyond news, and most of the answers are not in the model. They are in the architecture around it.

That architecture is the work. The newspaper is just where I learned to build it.

Have any thoughts? Send me a message >