MiroFish is a swarm intelligence prediction engine that builds high-fidelity parallel digital worlds populated by thousands of autonomous agents. Users feed it seed material — breaking news, policy drafts, financial signals, or even literary source text — and it returns a detailed prediction report plus a live interactive simulation environment. Currently at 49.3k GitHub stars / 7.2k forks (as of April 2026), making it one of the most-watched agentic AI repos. Latest release: V0.1.2 (March 7, 2026). Incubated by Shanda Group; simulation engine built on OASIS (CAMEL-AI's Open Agent Social Interaction Simulations framework).
How It Works
Five-stage pipeline:
- Graph Building — seed extraction, individual/collective memory injection, GraphRAG construction
- Environment Setup — entity relationship extraction, persona generation, agent configuration
- Simulation — dual-platform parallel simulation, auto-parsed prediction requirements, dynamic temporal memory updates
- Report Generation — ReportAgent with rich toolset for deep post-simulation interaction
- Deep Interaction — chat with any agent in the simulated world or query ReportAgent directly
Tech Stack
- Backend: Python 3.11–3.12, managed via
uv - Frontend: Vue (Node.js 18+)
- Agent memory: Zep Cloud (free tier sufficient for small simulations)
- LLM: Any OpenAI-compatible API; default config points to Alibaba Qwen-plus via Bailian Platform
- Deployment: Docker Compose or source; ports 3000 (frontend) / 5001 (backend)
- License: AGPL-3.0
Demonstrated Use Cases
- Public opinion simulation — modeled Wuhan University public opinion dynamics using BettaFish-generated seed report
- Literary prediction — simulated the lost ending of Dream of the Red Chamber from 80+ chapters of source text
- Upcoming: financial forecasting, political news prediction
Relevance to OpenClaw / Multi-Agent Work
MiroFish is architecturally adjacent to OpenClaw's multi-agent model but oriented toward simulation and prediction rather than task execution. Key differentiators: population-scale agents (thousands vs. a small coordinator+worker hierarchy), GraphRAG world-building from seed docs, and a ReportAgent layer for post-run analysis. OASIS as its simulation backbone is worth watching — could be a useful substrate or reference point for agent behavior modeling in OpenClaw contexts.