What the AI Actually Did Here
This system is part of our AI content suite. The message is simple: about 99% of what you see here was produced by AI.
This includes content creation and refinement, workflow ideation for this case study, website components and structure, translations, image generation, publishing, content management, CMS field population, and page assembly.
We treat AI as a tool, not magic. We closely monitor input and output quality, which is why this works.
AI does not create good results on its own. Garbage in leads to garbage out. Clear instructions, high-quality input, and structured configuration produce reliable output. The system amplifies good content; it does not invent strategy.
Step-by-Step: How the Pipeline Built This Page
AI refined the raw draft into publication-quality prose and analyzed the content, topics, and relationships. It designed the page architecture using predefined components, selected the appropriate hero, body, and FAQ blocks, and organized sections to create a clear narrative flow. SEO metadata—including titles, descriptions, and keywords—was generated automatically.
The system also derived FAQs from the content and likely reader questions, determined image placement, wrote image prompts and descriptions, generated the media, and applied rich formatting such as headings, lists, and emphasis. Where applicable, the content was translated into supported languages. Finally, an automated workflow published the page through the CMS.
Human reviewers then checked the result, made any final adjustments, and approved it.
The Problem: Traditional Content Publishing Is Manual and Slow
The problem we set out to solve is familiar to any content team.
Traditional content publishing is slow and manual. It relies on manual editing, rewriting, and tone adjustment; manual structuring of sections and page layout; manual component selection for each page; manual SEO research and optimization; manual decisions about image placement; and repetitive CMS entry—copy, paste, format, repeat.
Each article can take hours of human labor. Multiply that by dozens or hundreds of pages, and the bottleneck becomes obvious.
The Solution: A Multi-Agent, CMS-Native Pipeline
Our answer is a multi-agent AI content pipeline.
Instead of a single, overloaded agent trying to do everything in one prompt, we use multiple specialized agents, each with a clear responsibility.
The pipeline is multi-agent, with each agent handling a specific task in the chain; configuration-driven, with behavior controlled by configuration rather than hardcoded logic; CMS-native, with output shaped around real CMS components and fields; and fully automated, producing CMS-ready JSON with zero manual transformation.
From the user’s perspective, the flow is simple: provide raw content in almost any format, choose the page type and basic options, let the system analyze structure, intent, and assets, receive a refined, production-ready CMS payload, and review and approve.
The goal is a 99% AI-driven content pipeline, with humans in the loop only for final validation.
Inside the Pipeline: Six Specialized Agents
Each agent in the pipeline has one job.
The sanitizer cleans and normalizes inputs and loads the relevant configuration. The refiner polishes the prose to publication quality while respecting tone and constraints. The extractor analyzes structure and pulls out assets such as headings, quotes, links, media cues, and metadata candidates. The architect designs the page structure using the CMS component model and configuration rules. The SEO agent optimizes titles, descriptions, and other SEO fields. The builder assembles the final payload as CMS-ready JSON.
Each stage passes structured output to the next, rather than handing off a freeform block of text.
Behavior, Control, and JX-System Messages
Under the hood, the agents don’t behave in a static way. They dynamically adjust their behavior based on context and instructions while still remaining predictable for the user.
A small, well-structured prompt is enough to direct them to the right outcome. When you ask for a change, they focus on that change instead of rewriting everything.
We built this behavior using something we call JX-System messages. It’s a way to structure system-level instructions so agents can adapt to context without becoming chaotic.
We’ll explain JX-System messages in more detail in future articles. If you want to see how that works under the hood, make sure you subscribe to the mailing list so you don’t miss it.
