<- all tokdocs

MiroFish: An AI Agent Swarm That Simulates the Future

Watch on TikTok

View on TikTok ->

MiroFish is an open-source AI agent swarm engine built to predict the future through large-scale social simulation. It looks complex on the surface, but the core concept is straightforward: upload a seed event, let thousands of AI agents simulate how it plays out, and read the report. Built in just 10 days by a 20-year-old developer, it racked up over 33,000 GitHub stars and secured roughly $4 million in funding within 24 hours of launch.

How It Works

The pipeline has four stages:

1. Seed Material Upload

You start by uploading seed material -- a news article, a world event, a policy draft, anything you want to simulate forward.

2. Agent Generation via GraphRAG

MiroFish uses GraphRAG to read the input material and generate a set of agent personas relevant to the topic. Each agent gets its own profile, memory, and behavioral logic so they can all act independently within the simulation.

3. Multi-Agent Simulation via OASIS

Under the hood, MiroFish runs on OASIS (Open Agent Social Interaction Simulations), a simulation engine built by the CAMEL-AI team. OASIS is what lets the agents interact independently and evolve over time. It supports up to one million AI agents and 23 different social actions including following, commenting, and reposting.

Component Role
GraphRAG Reads seed material, generates agent personas
OASIS Simulation engine, supports up to 1M agents, 23 social actions
Report Agent Summarizes simulation results into actionable predictions

4. Report Generation

After the simulation completes, a report agent analyzes the results and summarizes patterns, outcomes, and emergent behaviors. You get a structured report you can use to inform decisions or explore potential futures.

Real-World Applications

The system has already been applied to practical use cases. One developer connected MiroFish to a Polymarket trading bot, simulated 2,847 digital humans before each trade, and reported $4,266 in profit over 338 trades. The approach represents a fundamentally different method of prediction -- using emergent behavior from simulated social dynamics rather than traditional statistical models.

Getting Started

MiroFish is completely open source. The entire system can be self-hosted with a few commands. For anyone new to multi-agent systems, the creator also offers a free series of lectures covering the foundations of building systems like this.

Key Takeaways

  • MiroFish uses AI agent swarms to simulate future scenarios from seed events
  • GraphRAG generates contextually appropriate agent personas automatically
  • The OASIS simulation engine supports up to 1 million interacting agents
  • The system was built in 10 days and hit 33,000+ GitHub stars
  • Completely open source and self-hostable
  • Already being used for real-world prediction tasks including trading

Resources

Published April 18, 2026. Writeup generated from a favorited TikTok.