For a Munich publisher, I developed a multi-agent system for content production that is running in production today: an end-to-end pipeline from topic to CMS. What mattered most was the architecture, far more than the prompts themselves.
Seven agents in one pipeline
The system consists of seven specialized agents. Each has its own role, its own context, and responsibility for one clearly defined step:
- Researcher – gathers sources and facts, creates the dossier and briefing
- Editor – writes the text
- Fact checker – checks every statement against the sources
- Tone of Voice – aligns language and brand with the magazine
- Formatter – puts the article into the CMS format
- SEO / GEO – optimizes it for search on Google and in AI answers
- Editor-in-chief – as an agent, makes the overall judgment before the editorial team approves
Optionally, a picture editor joins the pipeline. It selects images from the publisher's image database or generates images with AI, which are then labeled accordingly.
Together, the agents map an article's path from research to the CMS and fit into the publisher's existing workflows.
Boundaries, handoffs, and failure cases
Individual agents with good prompts are quick to build. The actual challenge lies in the interfaces between them. Where does one agent's task end? What does it hand over to the next one? What happens when a step delivers a weak result?
I treated these questions the way I would when designing a product. Agent boundaries, handoffs, and failure cases were defined before a single prompt was written. Each agent was given clearly defined constraints that ensure it stays within its role.
Quality assurance inside the pipeline
Quality control is built into the pipeline itself instead of being left to a human who corrects AI texts afterwards. If the quality is not sufficient, the piece goes through a quality loop before anyone gets to see it.
The editorial team still has the final say, because only what it has approved gets published. Its role thus shifts from fixing to deciding.
AI features or AI systems
In many projects, I see individual AI features being built where an AI system is actually needed. A feature answers a single request. A system takes over an entire workflow, with clear responsibilities, defined handoffs, and a plan for when something goes wrong.
Use in the day-to-day newsroom
The system is in production use in the day-to-day newsroom. Control remains with the editorial team, since nothing is published without three approvals. I developed it with Claude Code.
I describe how to build such a pipeline in Nine building rules for an AI pipeline at a publisher. The page about the AI editorial system shows how the system works in detail, and the example A newsroom day walks through an entire working day.
Want to build a digital product with a system?
Start a project