MiroFish is a multi-agent simulation platform that spins up thousands of AI agents, each with distinct personalities and persistent memory, to model collective behavior. You define a scenario in YAML: a market event, an election, a product launch, or a social dynamic. The agents interact, debate, and form emergent consensus. The output is a prediction with confidence intervals and a replay of how that consensus formed. The project went viral in early 2026 as a demonstration of swarm intelligence applied to social prediction. Demos showing stock market forecasts and public opinion simulation spread widely on X and HN. The technical approach borrows from agent-based modeling in economics and computational social science, but uses LLMs as the cognitive engine for each agent instead of fixed behavioral rules. MiroFish is genuinely weird and hard to categorize. Is it a prediction tool? A simulation platform? An AI research toy? Probably all three. It has practical applications in quantitative research and scenario planning, but also a delightful experimental quality that makes it fun to run for its own sake.

What the community says

MiroFish lit up social media when demos of its prediction accuracy circulated on X, with some users claiming it called market movements before they happened. Enthusiasts are drawn to its research potential and the sheer novelty of watching thousands of AI agents argue their way to consensus. Skeptics point out that validating prediction accuracy is hard and that the results depend heavily on the LLM being used. Some researchers note it as an interesting alternative to traditional agent-based models in social science.

See alternatives to MiroFish →