How Paul Iusztin built a multi-agent workflow to research, write and edit LinkedIn posts

The problem

Paul says people try to fix same-sounding AI posts by writing “better” prompts. In his view, a single prompt can’t reliably enforce these six things:

His alternative order of work is “Research first. Write second. Evaluate third. Edit until it improves.”

Research. A deep research agent takes a topic and researches it autonomously. It uses grounded web search, YouTube video analysis and follow-up queries, then compiles source-backed research. The output is saved as research.md.
Write in a loop. A LinkedIn writing workflow takes research.md plus guideline.md and runs “Generate → review → edit → repeat”.
Check each pass. Every pass checks for weak hooks, AI slop, structure issues, voice drift and missing grounding.
Keep every version. “Every intermediate version is saved, so you can inspect exactly what changed.”
Observe and evaluate. Opik traces LLM calls, tool calls, latency, cost, inputs and outputs, and evaluation results. His reasoning: “Because if you can’t observe your agents, you can’t improve them.”

Claude Code orchestrates the system. FastMCP exposes the tools, and Gemini handles grounded search and video analysis.

His summary: “Good AI writing is not just prompting…” and “It’s also system design.”

None stated. The post includes no example posts, engagement figures, cost or latency numbers, or before-and-after comparisons.

The research, write, evaluate, edit order applies to any distribution draft, not only LinkedIn posts.
Saving every intermediate version gives a human reviewer a record of what changed before anything ships.
Checks like “voice drift” and “missing grounding” work as written review criteria for our own drafts.

The post sells a course. Paul says the workflow is now “a full professional course lesson inside our Agentic AI Engineering course” and “Available in our free preview.” The post doesn’t state pricing for the rest of the course.
The post says the course lesson covers the architecture, the MCP setup, the evaluator-optimizer loops, the observability layer, the eval pipeline and the full implementation. The post itself gives only an outline.
The post says the system “was originally built as a live workshop for AI Engineer.”
The post names and tags Opik (by Comet) and FastMCP (by Prefect). It doesn’t state any relationship with those companies.
The post was built by more than one person (“we built”). The source text doesn’t say who else was involved.
The source text doesn’t give a post date. The comments show “5mo”.
The comments are reader reactions and aren’t used as evidence here.
X: Unverified not checked

Want agents like these working your distribution? See the Agent Stack Teardown. Get new teardowns weekly: subscribe (double opt-in).