Simulate future scenarios by seeding thousands of AI agents with news or fiction
Upload an interesting news article or novel manuscript, and thousands of artificial intelligences inhabit it, conversing and acting with one another. This process resembles a laboratory that reproduces real-world social interactions in a digital space. When you inject a desired scenario, the system returns the resulting future trends in a reporting format.
Prediction experiments starting with seed data
This tool builds a virtual world based on provided information. When a user uploads 'seed' materials such as news or reports and describes the prediction target in natural language, the system generates agents with independent personalities and memories. These agents interact with each other, undergo social evolution, and infer future trajectories in the process. Users can inject variables from a 'god's eye' perspective to simulate how specific outcomes change.
Graph-based memory and agent composition
Internally, the system extracts individual and group memories from seed information and organizes them into a graph structure. Based on this graph, it sets the personality and behavior logic for each agent, completing the environment through relationship extraction. As the simulation progresses, memories are updated over time, and simulations run in parallel on two platforms. In the final stage, a report agent equipped with a toolset interacts with the post-simulation environment to generate a detailed report.
Installation and execution environment
Source code distribution is recommended, requiring Node.js 18 or higher, Python 3.11 to 3.12, and the uv packaging manager. You must enter LLM API keys and Zep Cloud API keys in the environment variable file, and you can use LLM APIs that support the OpenAI SDK format. The official documentation recommends the Alibaba Qwen-plus model and advises trying simulations of fewer than 40 rounds for initial testing. Dependency installation and service execution are handled by the npm run setup:all and npm run dev commands, with the frontend running on port 3000 and the backend on port 5001.
Costs and limitations to check before use
Because this tool performs large-scale agent simulations, LLM API usage can be significant. The README explicitly states to reduce the number of rounds for testing in the initial stages, indicating that a gradual approach is needed rather than unlimited experimentation. It also mentions that Zep Cloud's free allocation is sufficient for simple use, suggesting that additional cost burdens may be low in the early stages. Currently, the demo focuses on public opinion simulation and novel ending prediction, while financial or political news prediction is noted as planned for future updates.
