MiroFish: The Multi-Agent AI Engine That Simulates the Future

What if you could rehearse the future before it happens? That's the promise behind MiroFish, an open-source multi-agent AI prediction engine that constructs parallel digital worlds populated by thousands of autonomous AI agents—each with its own personality, memory, and behavioral logic.

What Is MiroFish?

MiroFish extracts seed information from real-world sources—news articles, policy drafts, financial signals—and uses it to build high-fidelity digital simulations. Within these simulations, AI agents interact, form opinions, and evolve socially, producing emergent outcomes that serve as detailed forecasts.

The core idea is elegantly simple: let the future rehearse in a digital sandbox, then make decisions after a hundred simulations.

Users upload seed materials, describe their prediction needs in natural language, and receive detailed forecast reports alongside an interactive sandbox for deep exploration.

How It Works: A Five-Stage Pipeline

MiroFish follows a structured simulation pipeline:

1. Graph Building — Seed data is extracted and injected into knowledge graphs using GraphRAG, establishing the informational foundation for the simulation.
2. Environment Setup — Entities are extracted, characters are generated with distinct personalities, and simulation parameters are configured.
3. Simulation Launch — Thousands of agents run in parallel across dual-platform simulation environments, with automatic prediction interpretation and time-series memory updates.
4. Report Generation — A specialized ReportAgent synthesizes findings using multiple analytical tools into comprehensive forecast reports.
5. Deep Interaction — Users can converse directly with simulated entities or analysis agents, injecting dynamic variables to test alternative futures from a "god's perspective."

Key Features

• Knowledge Graph Construction — Builds structured representations of real-world scenarios with GraphRAG integration
• Autonomous Agent Simulation — Deploys thousands of agents with independent personalities, long-term memory, and behavioral reasoning
• Dual-Platform Parallel Processing — Runs multiple simulation scenarios simultaneously for faster results
• Interactive Analysis — Supports natural language conversations with simulated entities and report agents
• Dynamic Variable Injection — Adjust simulation parameters in real-time to explore alternative outcomes
• Individual and Collective Memory — Agents maintain both personal memory and shared group knowledge

Real-World Use Cases

MiroFish shines across a range of applications:

Macro-Level Predictions

• Policy testing — Simulate how a new regulation might affect public behavior before implementation
• Public relations scenario planning — Model how a crisis might unfold across social networks
• Decision-maker simulation labs — Test strategic decisions in a risk-free environment

Micro-Level Explorations

• Narrative prediction — The team demonstrated MiroFish predicting plausible endings for classic literature based on 200,000+ characters of source material
• Creative exploration — Writers and content creators can simulate how storylines might evolve

Domain-Specific Analysis

• Public opinion forecasting — Successfully validated against real-world Wuhan University public opinion scenarios
• Financial signal analysis — Model market reactions to economic events
• Social trend prediction — Understand how cultural shifts might propagate through populations

Technical Stack

MiroFish is built on a modern, accessible stack:

• Frontend: Vue.js
• Backend: Python (3.11–3.12)
• Simulation Engine: Powered by OASIS (CAMEL-AI framework)
• Memory System: Zep Cloud integration
• LLM Integration: Any OpenAI SDK-compatible API (Alibaba Qwen-Plus recommended)
• Deployment: Source code or Docker, with frontend on port 3000 and backend API on port 5001

Setup is straightforward—configure your .env file with LLM API credentials, run npm run setup:all, then npm run dev.

Why It Matters

Traditional forecasting relies on statistical models that struggle with the complexity of human behavior. MiroFish takes a fundamentally different approach: instead of modeling trends, it models people—thousands of them—and lets their interactions produce emergent outcomes.

This agent-based approach captures dynamics that statistical methods miss: social influence, opinion cascading, behavioral feedback loops, and the kind of nonlinear effects that make real-world prediction so difficult.

With over 36,000 GitHub stars and active development backed by Shanda Group, MiroFish represents a significant step toward making multi-agent simulation accessible to researchers, analysts, and decision-makers.

Get Started

MiroFish is open-source under the AGPL-3.0 license. Explore the project on GitHub, join the community on Discord, and start building your own digital prediction sandboxes.

---

Interested in AI-powered generation tools? Hypereal AI provides unified API access to the best AI models for image, video, and audio generation.