AIMM Session — August 7, 2025: The Multi-LLM Debate Workflow

“This is our planting season. Get in early while it’s compounding — because if you get in after the compounding, it’s too late.” — Lou, on AI adoption timing

30-Second Summary

Dense session. Opened with a real conversation about AI’s economic disruption — from warehouse robots to Hollywood to your CBT coach — and closed with a live N8N build: a multi-LLM debate workflow where GPT, Claude, and Grok argue a question in a shared memory space, and a fourth model synthesizes the winner. Lou also previewed a transcript-to-content pipeline that turns coaching calls into newsletter essays automatically.

1. The LLM Subscription Math Problem

You might be paying for three subscriptions when one API key could do the job.

Lou opened with a question that’s probably been nagging at anyone running multiple paid AI accounts: is it worth it? Three subscriptions (ChatGPT Plus, Claude Pro, Grok) add up to roughly $1,000 CAD/year — and none of them show you how much you’re actually using.

Two alternatives stood out:

• Abacus.ai ($10/month) — Donald had just gifted it to his brother. Access to all major LLMs, strong research capability, agent support. Lou called it “well worth it.”
• Open Web UI — Self-hosted, open source, data stays on your server. Lou’s preferred option for anyone who cares about data ownership. Run multiple models concurrently and even merge their outputs.

The real point: the consolidation problem matters more than the cost. Having your prompts, conversations, and context scattered across four platforms is a workflow tax that compounds over time.

2. AI & The Job Market: Two Conferences, Two Realities

Ri Ca attended two conferences in the same two-week stretch. The contrast was striking:

• Business conference: AI will kill jobs. Solution? Bread and circuses — socialist safety nets to avoid revolution.
• SAS software conference: AI will create more work, not eliminate it. Your job is safe; you’ll just oversee the outputs.

Lou’s take: East/West automation direction matters. Japan and China automate from the bottom up (factories, warehouses). North America automates from the top down (coding, finance, law, creative work).

“The people I worry about most are the ones who can’t easily replace their jobs — they haven’t been exposed to the training.”

Don Back’s historical frame:

“The whole concept of paying someone for labor came from attracting people out of itinerant farming. If that’s no longer available, how do we move back to a situation where people can satisfy their addiction to food, clothing, and shelter?”

Donald Kihenja added the global lens: “Even people in Kenya who don’t have computers — entire generations will just miss out.”

3. The CBT Coach Parable

Lou had signed up for a productized CBT (cognitive behavioral therapy) program — the therapy itself is now a self-guided product; the coach’s job is to check your implementation. That productization had already reduced hiring for coaches.

He asked her directly: What are you doing to prepare yourself for the next two to three years?

Her answer: the company will rewrite their programs with AI, so she’ll be fine.

Lou’s counter: when they do that, they won’t need as many of you. Her fallback: the personal touch. Lou’s counter: “The personal touch is a distinguishing factor, but it’s not enough for people to work with you over a productized AI that has your curriculum baked in.”

The practical takeaway:

“You’re gonna be on your own, trying to compete with AI and companies that use AI. So you need to start asking yourself now: what can I do to deliver better client experience using AI? What can I automate so I can offer more of this personal connection?“

4. OpenAI Agent Builder: Skip It (For Now)

Lou’s verdict: version 1 of something N8N already does better. Locked to the OpenAI ecosystem. Feels like a revival of deprecated products (GPTs, Assistants, Plugins) with a drag-and-drop front end.

Recommended hierarchy:

1. N8N (self-hosted, most powerful, steepest curve)
2. Make.com (middle ground)
3. Zapier (reliable, simple, expensive)
4. Agent Builder (wait a year)

5. The Multi-LLM Debate Workflow (Live N8N Build)

The main event: teaching multiple AIs to argue, then synthesize.

The architecture Lou built:

• Three AI Agent nodes in N8N, each using a different LLM via OpenRouter (one API key, access to all models)
• All three agents share a single memory node — the “whiteboard” approach
• Each agent reads the full conversation history and adds its perspective
• A fourth “arbiter” agent (DeepSeek, on the free tier) reviews everything and produces a synthesized final answer

The key insight: connecting all three agents to the same memory store instead of passing outputs manually between nodes. The agents automatically see each other’s contributions as the conversation builds.

Lou tested it live with: “What’s the best RAG architecture for a knowledge base containing a Canadian training course on copywriting?”

Where this goes next:

• Add a loop node so agents debate across multiple rounds
• Replace the static agent list with a dynamic JSON config
• Connect via MCP so you can trigger the entire debate from any chatbot with one command
• Swap models for roles: researcher → content planner → writer → editor

“Imagine if you could send that JSON file to the workflow via MCP — created by Claude, parameterizing what experts you need and what models to use. You would have a completely dynamic environment.”

Hot Takes

• “We are all the CEOs of our own careers — we need to act accordingly.” — Don Back
• “Information is essentially free. Transformation and implementation are the valuable piece.” — Don Back
• “If you take a practitioner’s curriculum and put it into an AI, it would be as good, if not better, than any of their practitioners.” — Lou

Tools Mentioned

Tool	What It Is
Abacus.ai	$10/mo multi-LLM platform
Open Web UI	Self-hosted multi-model chat UI
OpenRouter	Single API for all major LLMs
N8N	Open-source workflow automation
DeepSeek v1 (free)	Open LLM via OpenRouter — used as arbiter

Try This Before Next Session

Build the debate workflow in N8N — from scratch.

1. Set up a free OpenRouter account and grab your API key
2. Create a new N8N workflow with a Chat Trigger node
3. Add three AI Agent nodes — assign each a different model via OpenRouter
4. Connect all three agents to a single shared Simple Memory node
5. Add a fourth AI Agent as the “arbiter” — system prompt: “Review the conversation history, assume the role of a top 0.1% expert in the relevant domain, evaluate each input, and produce a comprehensive final response.”
6. Run it with a real question you’ve been wrestling with

Lou’s estimate: one to two hours if it’s your first N8N build. The goal: by the end of the month, this workflow lives on your MCP server and you can trigger a multi-model debate from any chatbot with one command.