666ghj/mirofish

MiroFish is a multi-agent simulation engine that tries to predict future outcomes by creating a digital world populated with thousands of AI characters who interact with one another. The problem it aims to solve is that traditional forecasting methods rely on statistics and historical patterns, which struggle with complex social phenomena where the behavior of individuals affects the behavior of others in cascading ways. MiroFish approaches this differently: you feed it real-world seed material, such as a news article, a policy document, or financial data, and it builds a simulated environment with AI agents representing people or entities, each given their own personality, memory, and decision-making logic. These agents then interact freely over many simulated time steps, and the resulting collective behavior is used to generate a prediction report. The workflow has four stages: first it builds a knowledge graph from the seed material, then it creates the simulated environment and configures agent personalities, then it runs the simulation on two parallel platforms, and finally a dedicated reporting agent synthesizes the interactions into a readable prediction. You can also directly chat with any agent in the finished simulation to explore specific scenarios. The tech stack uses Python on the backend and Node.js for the frontend, with support for LLM APIs in the OpenAI SDK format (the documentation recommends Alibaba's Qwen-plus model). Agent memory is handled by Zep Cloud. The application runs either from source code or via Docker Compose. The project is aimed at use cases including financial forecasting, public opinion analysis, political scenario modeling, and even more creative applications like predicting how a fictional story might end. Topics listed include multi-agent simulation, knowledge graphs, and financial forecasting.