Most multi-agent demos are a chat loop with a system prompt. This session is a real system that has been shipping for months.

powerplatformninja.com is my technical blog: 2,000-word deep dives on Power Platform and Azure AI engineering, for people who already ship real solutions. Every post follows a house style — a contents list, five to seven sections, a “What to watch carefully” critical-reading section, a “My take” close, cited sources — and that style is exactly what makes it automatable. A crew of seventeen agents orchestrated with Microsoft Agent Framework on Azure AI Foundry now researches, outlines, writes, fact-checks, illustrates and pushes each article into WordPress. Nothing auto-publishes: the crew’s job ends at a draft plus a review report, and I press Publish. Then the system learns from what I changed.

We’ll walk the actual graph. Fan-out/fan-in across three research scouts and two independent validators. Two bounded feedback loops — re-research and revise — and why there is deliberately no third. Conditional entry points that resume a failed run from saved research instead of paying for it twice. A human-in-the-loop approval split across two separate graphs, so no worker is ever held open waiting for a person.

The rule underneath all of it: agents produce typed judgements; code makes every routing decision. Every agent is bound to a structured output contract, and the gates between them are plain, testable logic.

Then the rare part: a supervised self-learning loop. The system diffs its own draft against what I actually published, scores the edit rate, clusters what recurs, and proposes configuration changes — but replays every candidate rule against everything already published first. A rule that fires on text the human shipped is a false positive by construction.

Finally, the production scars: the SDK contract that fails only on real infrastructure, the token meter attached at the wrong seam, and the bug that silently unsubscribed a live recipient.