I wanted to understand how multi-agent systems actually work — not from a tutorial, but by wiring the pieces together myself. So I built a medieval townhall simulation where three AI characters (a mayor with a French accent, a judge, and a villager) debate topics in a medieval townhall. They can all hear each other in real time and can all contribute to the conversation.
This project covers how to implement asynchronous agent communication, how to pass a shifting context window, how to store and retrieve memory using vector search (cosine similarity search) and finally lessons I learned on how to get the best outputs out of your agents based on the Multi-Agent Debate (MAD) framework.
Find the repository here.
The Architecture
How do we get asynchronous communication?
