A project on GitHub just blew past 35,000 stars in a matter of weeks:
What MiroFish Is and How It Works
MiroFish is a swarm intelligence engine that starts from real-world "seeds" โ news articles, policy documents, financial reports, even novels โ and builds a parallel digital world on top of them. Inside this environment, thousands of agents with distinct personalities, long-term memory, and behavioral logic interact freely.
The pipeline has five stages:
- Graph construction: seed information extraction, individual and collective memory injection, GraphRAG building
- Environment setup: entity relationship extraction, persona generation, simulation parameter injection via configuration agents
- Simulation: dual-platform parallel execution, automatic prediction request parsing, dynamic temporal memory updates
- Report generation: a dedicated ReportAgent interacts with the post-simulation environment using a rich toolset
- Deep interaction: you can talk to any agent in the simulation or to the ReportAgent directly
Where a Swarm Engine Actually Matters
Two demonstrated use cases, very different from each other. The first is serious: public opinion prediction. Wuhan University used MiroFish to simulate public reactions to breaking news events and generate predictive reports. The second is more creative: they fed the engine the surviving 80 chapters of Dream of the Red Chamber (็บขๆฅผๆขฆ) and had it predict the lost ending of the classic Chinese novel.
This is where it gets interesting. MiroFish is not a traditional statistical forecasting system. It is an emergent approach: let collective agent behavior surface patterns and trends, exactly like biological swarm intelligence. The output is closer to an "automated thought experiment" than a regression model.
Practical Limits and Considerations
Before getting too excited, some notes. Token consumption is high โ the README itself warns to start with fewer than 40 simulation rounds. Every agent calls an LLM each turn, so costs scale fast. Setup requires Node.js 18+, Python 3.11โ3.12, and uv as the package manager. Docker deployment is also available and simplifies things.
The simulation engine is built on OASIS by CAMEL-AI, an open-source multi-agent simulation framework. MiroFish adds the real-data injection layer, the reporting system, and the interactive interface on top. The project is backed by Shanda Group and released under the AGPL-3.0 license.
One thing I appreciate: the ability to interact with individual agents after the simulation. It is not just "read the report." You can step into the simulated world and ask an agent directly about its reasoning.
FAQ
Does MiroFish only work with Chinese LLMs? No. It supports any API compatible with the OpenAI SDK format. You can use GPT-4, Claude, Qwen, or local models served through a compatible server.
How much does it cost to run a simulation? It depends on the number of agents and rounds. The project recommends starting with fewer than 40 rounds to keep costs manageable. Each agent generates at least one LLM call per turn.
Can I use it for real financial predictions? The project lists financial forecasting among its use cases, but it is an experimental tool. Do not use it for investment decisions without independent validation.