Generate Consensus Answers with Multiple AI Models & Peer Review System

Description

AI Council: Multi-Model Consensus with Peer Review

Inspired by Andrej Karpathy's LLM Council, but rebuilt in n8n.

This workflow creates a "council" of AI models that independently answer your question, then peer-review each other's responses before a final arbiter synthesizes the best answer.

Who is this for?

If you want to prepare for an upcoming meeting with different people and prep for their different views
find any "blind spots" in your view on a certain subject
Researchers wanting more robust AI-generated answers
Developers exploring multi-model architectures
Anyone seeking higher-quality responses through AI consensus, potentially with faster/cheaper models.
Teams evaluating different LLM capabilities side-by-side

How it works

Ask a Question — Submit your query via the Chat Trigger
Individual Answers — Four different models (Gemini, Llama, Gemma, Mistral) independently generate responses
Peer Review — Each model reviews ALL answers, identifying pros, cons, and overall assessment
Final Synthesis — DeepSeek R1 analyzes all peer reviews and produces a refined, consensus-based final answer

Setup Instructions

Prerequisites
Access to an LLM (e.g. OpenRouter account with API credits)

Steps
Create OpenRouter credentials in n8n:
Go to Settings → Credentials → Add Credential
Select "OpenRouter" and paste your API key
Connect all model nodes to your OpenRouter credential. In this example I used Gemini, Llama, Gemma, Mistral and Deepseek, but you can use whatever you want. You can also use the same models, but change their parameters. Play around to find out what suits you best.
Activate the workflow and open the Chat interface to test

Customization Ideas

You can add as many answer and review models as you want. Do note that each AI node is executed in series, so each will add to the total duration.
Swap models via OpenRouter's model selector (e.g., use Claude, GPT-4, etc.)
Adjust the peer review prompt to represent a certain persona or with domain-specific evaluation criteria
Add memory nodes for multi-turn conversations
Connect to Slack/Discord instead of the Chat Trigger