MiroFish: Simulate Thousands of Agents with News and Novels to Predict the Future

When seed information such as news, policy drafts, or financial signals is input, a parallel digital world is automatically generated where thousands of intelligent agents interact with independent personalities and long-term memory. Users simply describe their prediction requirements in natural language and receive detailed prediction reports derived from the simulated environment along with an interactive digital world.

Through GraphRAG-based graph construction and entity relationship extraction, entity-specific personas are generated, and dynamic temporal memory is updated via dual-platform parallel simulation. After the simulation ends, users can engage in in-depth conversations about the results through ReportAgent, or directly interact with specific agents within the virtual world to inject variables and precisely infer future trajectories.

Unlike existing prediction models, this approach aims to surpass the limitations of traditional prediction by capturing collective emergence arising from individual interactions. It is useful for decision-makers and creators who want to validate various scenarios in a zero-risk environment, from policy rehearsals to novel ending inference. Developed based on the OASIS engine with support from Shanda Group, it is released under the AGPL-3.0 license.