I've been building agent workflows in n8n for a while now, mostly to solve my own problems. This one started because I kept doing the same thing manually: paste a document into a few different AI chats, ask each one to poke holes in it from a different angle, then try to mentally merge the answers myself. That's slow, and I always forget what the first chat said by the time I've read the third.

So I built a workflow that does it for me. A panel of AI personas reads a document, argues about it across a few rounds, and hands back a structured report. I've been calling it the AI Round Table.

What It Actually Does

You upload a document, PDF, Markdown, or plain text, through a form. You can also add a line of context about what kind of analysis you want or what decision it's feeding. The workflow takes it from there.

First it reads the document and picks a panel. There are four fixed personas that show up every time:

• Maya, the skeptic, who probes claims for unsupported assumptions
• Devon, the domain expert, who checks technical and subject matter accuracy
• Priya, the feasibility pragmatist, who weighs real-world constraints and execution risk
• Jules, the devil's advocate, who argues the strongest opposing case

On top of those four, an AI call reads your specific document and proposes two or three more personas tailored to it. Feed it a business plan and you might get a numbers-focused CFO type. Feed it a technical design doc and you'll probably get someone stress-testing the architecture. The panel lands at five to seven people depending on how much the document actually calls for.

The Discussion Itself

Once the panel is set, they go through rounds. Round one, everyone gives an independent take on the document from their assigned angle. From round two on, each persona reacts to what the others said, refines their position, pushes back, or introduces something new.

This isn't unbounded. It's capped at five rounds, and there's a moderator agent watching the discussion starting in round two. If a round mostly repeats points already made, the moderator counts it as a zero-new-point round. Two of those in a row and the discussion stops early. No point paying for agents to talk past each other.

If the transcript grows past a certain length, a distillation step compresses it down into a dense summary before the next round starts, so the context doesn't balloon and the later agents aren't rereading a wall of near-duplicate text.

What You Get Back

At the end, a synthesis step reads the whole discussion and produces a report with:

• The panel roster, names and descriptions
• Each persona's final position, plus whether and why they shifted from their round-one take
• An overall analysis covering where the panel agreed, where it diverged, and what that pattern suggests
• Arguments for and against, pulled from the actual discussion
• A feasibility assessment, when the document is the kind of thing that warrants one
• A risks section, same condition
• A full round-by-round transcript as an appendix, in case you want to see how someone got to their final position

The report lands in a Google Doc, and I get a Telegram ping when it's done so I don't have to sit there watching the execution log.

Why I Built It This Way

The moderator and the round cap exist because I didn't want a workflow that burns tokens for the sake of looking thorough. Five agents having a real disagreement for two rounds is more useful than seven agents restating themselves five times. Bounding it and giving it a way to stop itself was the point, not an afterthought.

The dynamic personas were the other piece I cared about. A fixed panel of four is fine, but it doesn't know anything about what you fed it. Having the workflow read the document first and pick a couple of personas suited to it means the panel actually engages with what's in front of it instead of running the same four questions on everything from a marketing brief to a technical spec.

It's on Gumroad

I packaged this up as an n8n workflow template and put it on Gumroad. If you want the whole thing (intake form, persona generation, the round loop, moderator, distillation, synthesis, all of it) it's here:

AI Round-Table: Multi-Agent Debate & Analysis Workflow

Import it, point it at your own Gemini (or model of your choice) and Google Docs credentials, and you've got a panel of AI personas ready to argue about whatever you throw at them.