MiroFish Builds a Fake World Out of Agents and Asks It What Happens Next
MiroFish is at 73,100 stars and was up another 524 on GitHub Trending today, AGPL-3.0, at https://github.com/666ghj/MiroFish. It describes itself as a simple and universal swarm intelligence engine that predicts anything, which sounds like nonsense until you look at what it actually does, at which point it becomes a very specific and slightly unnerving thing.
You give it seed material — reports, news, a story — and a question about what happens next. It extracts a knowledge graph from the seeds with GraphRAG, generates entities and personas from that graph, populates a simulated environment with thousands of agents carrying distinct personalities, runs them in parallel across two platforms, and then has a separate set of reporting agents write up what happened. You can chat with the simulated entities afterward and ask them why they did what they did. Target use cases are finance, public opinion and narrative analysis, which is to say: what will the market do, what will people say, how does this story spread.
The engine underneath is OASIS from the CAMEL-AI framework, which has been the serious open-source social-simulation stack for a while now. The interesting engineering choice is the pipeline shape: graph construction, environment setup, parallel simulation, then a reporting layer made of agents rather than a dashboard. The output is a written analysis, not a probability. That is honest about what this can and cannot do — nobody should be reading a calibrated forecast out of a thousand LLM personas — but it also means the thing that comes out the end is exactly as checkable as an essay, which is not very.
Worth being blunt about the epistemics. A simulation of a thousand agents reproduces the model's priors about how people behave, at scale, with a lot of confident prose attached. When it is right it is because the seed documents carried the signal, and when it is wrong it will be wrong fluently. Treat it as a structured brainstorming machine for scenarios you had not considered, which is genuinely valuable, and not as a forecaster. [The agent research tools that have held up](https://clauday.com/article/b8841d6b-71bb-44de-b5e1-3e71c6d022ca) are the ones that put critics and verification in the loop, and a persona swarm has neither.
The star count deserves a caveat too. 73,100 stars on a project created in late November 2025, with backing from Shanda Group, is an extraordinary curve, and the last real commit was September 3. The code and the architecture are real; the growth pattern is the kind you should look at twice.
← Back to all articles
You give it seed material — reports, news, a story — and a question about what happens next. It extracts a knowledge graph from the seeds with GraphRAG, generates entities and personas from that graph, populates a simulated environment with thousands of agents carrying distinct personalities, runs them in parallel across two platforms, and then has a separate set of reporting agents write up what happened. You can chat with the simulated entities afterward and ask them why they did what they did. Target use cases are finance, public opinion and narrative analysis, which is to say: what will the market do, what will people say, how does this story spread.
The engine underneath is OASIS from the CAMEL-AI framework, which has been the serious open-source social-simulation stack for a while now. The interesting engineering choice is the pipeline shape: graph construction, environment setup, parallel simulation, then a reporting layer made of agents rather than a dashboard. The output is a written analysis, not a probability. That is honest about what this can and cannot do — nobody should be reading a calibrated forecast out of a thousand LLM personas — but it also means the thing that comes out the end is exactly as checkable as an essay, which is not very.
Worth being blunt about the epistemics. A simulation of a thousand agents reproduces the model's priors about how people behave, at scale, with a lot of confident prose attached. When it is right it is because the seed documents carried the signal, and when it is wrong it will be wrong fluently. Treat it as a structured brainstorming machine for scenarios you had not considered, which is genuinely valuable, and not as a forecaster. [The agent research tools that have held up](https://clauday.com/article/b8841d6b-71bb-44de-b5e1-3e71c6d022ca) are the ones that put critics and verification in the loop, and a persona swarm has neither.
The star count deserves a caveat too. 73,100 stars on a project created in late November 2025, with backing from Shanda Group, is an extraordinary curve, and the last real commit was September 3. The code and the architecture are real; the growth pattern is the kind you should look at twice.
Comments