If you could clone a small society, let them freely discuss something — and then watch how their public opinion evolves — you’d basically have a “future predictor.”
This isn’t science fiction. MiroFish is doing exactly that.
What Is MiroFish?
MiroFish is an open-source multi-agent social simulation prediction engine that’s already gathered over 16,000 stars on GitHub. Backed by Shanda Group.
Its core concept is pretty intuitive:
- Give it a real event (news, policy, earnings report, anything goes)
- It automatically generates thousands of AI Agents, each with independent personalities, memories, and behavior logic
- Let these Agents interact freely on simulated Twitter and Reddit
- Observe how group public opinion evolves
- Output a prediction report
Simply put: it builds a parallel digital world and lets you observe “if this happened, how would people react?”
How Does It Work? Five Stages
Stage 1: Knowledge Graph Construction
You upload raw data — a news article, an earnings report, a policy draft — and MiroFish uses GraphRAG technology to break it down into a knowledge graph. It’s not just storing text; it understands the relationships between entities.
Stage 2: Agent Generation
The system automatically generates a large number of AI Agents. Each Agent has its own:
- Personality traits (optimistic/pessimistic, risk appetite, professional background)
- Long-term memory (managed through Zep Cloud)
- Behavior logic (not acting from a script — they respond autonomously based on their personality)
Stage 3: Community Simulation
These Agents are placed into simulated social platforms, free to post, comment, retweet, and argue. You can “inject variables” during the simulation — like a god’s-eye view, drop in a new piece of news, and watch how group reactions change.
Stage 4: Report Generation
ReportAgent analyzes all the simulation data and outputs a structured prediction report. It’s not simple statistics — it analyzes turning points in public opinion, key influencing factors, and sentiment trends.
Stage 5: Deep Interaction
You can directly chat with Agents in the simulation and ask them “Why do you think that way?” — the answers you get are based on their complete experiences and memories within the simulation.
Why Does This Matter for Investors?
If you trade or invest, you definitely care about “market sentiment.”
Traditional methods look at the Fear & Greed Index, social media volume, and fund flows. But these are all things that “already happened” — you’re seeing the result, not the process.
MiroFish’s approach is different: it simulates the formation process of sentiment.
Imagine these scenarios:
- Fed announces a 50 basis point rate hike → simulate how retail investors, institutions, and analysts each react, and which direction public opinion converges
- A crypto exchange gets hacked → simulate investors with different risk appetites, predict where capital will flow
- New regulatory legislation passes → simulate different responses from communities in various countries
This isn’t replacing technical analysis — it’s adding another dimension: “group behavior prediction.”
Tech Stack Overview
MiroFish’s tech stack is pretty developer-friendly:
- Frontend: Vue.js (accounts for 41% of the code)
- Backend: Python 3.11-3.12 (accounts for 58%)
- Simulation Engine: CAMEL-AI’s OASIS framework
- Knowledge Management: GraphRAG
- Memory System: Zep Cloud (the free plan is enough)
- LLM: Supports any OpenAI SDK-compatible model, defaults to Alibaba’s Qwen
Two deployment options: source code installation or Docker Compose one-click deployment. Low barrier to entry.
My Take: Why This Project Is Worth Watching
1. Multi-Agent Isn’t Just Chatting
Most “multi-Agent” projects on the market just have a few AIs talk to each other. MiroFish’s scope is different — it simulates social behavior. Agents don’t just converse; they form groups, create herd effects, and develop opinion leaders. This is closer to the real world.
2. Injectable Variables in Simulation
You can toss new information into the simulation anytime (“suddenly announces a rate cut”) and watch group reactions change in real time. This “god’s-eye view” design is incredibly powerful because you can’t do controlled experiments like this in the real world.
3. Prediction ≠ Prophecy
MiroFish won’t tell you whether BTC will go up or down tomorrow. What it tells you is: “In this scenario, which direction is group sentiment most likely to move toward.” This is probabilistic thinking, not deterministic prediction — when used right, it’s more valuable than any “price call.”
4. Open Source + AGPL-3.0
Fully open source — you can run it and modify it yourself. AGPL license means derivative versions must also be open source, but there’s no issue for personal research and internal use.
Limitations and Risks
Fair is fair — let’s talk about the drawbacks:
- LLM costs: Simulating thousands of Agents requires a massive number of API calls, which adds up
- Simulation ≠ Reality: No matter how realistic AI Agents are, they’re not real people. Group behavior complexity far exceeds what models can cover
- Bias risk: Agent personality settings are based on LLM training data, which may have systemic biases
- Project maturity: Despite high star counts, it’s still in rapid iteration — be cautious when using in production
Conclusion: Worth Watching, but Stay Rational
MiroFish represents an interesting direction: using group simulation for prediction instead of using historical data for backtesting.
These aren’t mutually exclusive. The strongest prediction systems should combine technical analysis + fund flows + group behavior simulation.
If you’re a developer, it’s worth spending an afternoon getting it running. If you’re an investor, just understanding the concept is enough — it’ll change how you think about “market sentiment.”
MiroFish GitHub: github.com/666ghj/MiroFish
This is one of our team’s research notes on multi-agent systems. We’re also using AI Agents for trading strategy development and automation — if you’re interested in this field, feel free to follow our updates.