A Multi-Agent Pipeline That Researches, Writes and Publishes Blogs

The situation

A good article is several jobs: finding a topic worth covering, researching it, planning the outline, writing, optimising for search, creating images, reviewing and publishing.

One prompt to one model does all of them badly. We wanted the quality of a small editorial team, running on a schedule.

What we built

A backend that runs each article through a fixed pipeline, with one agent responsible for each stage.

• Manager: decides whether a topic should run, be skipped or be rejected as a duplicate of something already covered.
• Researcher: gathers current sources through web search.
• Strategist: sets the search intent, the target length and the outline.
• Writer: drafts the article to that outline.
• SEO: optimises the draft for search.
• Designer: generates and compresses the images.
• Marketing & structured data: prepares the structured data that goes with the article.
• Reviewer: scores the article and, if it falls short, sends it back to the writer.
• Publisher: sends the finished article to WordPress or to the website’s CMS by webhook.

Controls

Topics are discovered through web search for each project’s industry, and near-duplicate ideas are dropped before they become drafts. Competitor brand names are kept out of titles.

Every agent’s prompt lives in a versioned file, every run and agent step is logged, and a failed stage can be retried without starting the article again.

The result

One platform serves several website projects, each with its own topics and publishing destination. Articles arrive in the CMS already reviewed, with images and structured data attached.