AI Tools

MiroFish: The AI That Simulates Entire Societies to Predict the Future (36K Stars in Days)

Feed it a news article, a policy draft, or market data. It creates a parallel digital world populated by thousands of AI agents — each with their own personality, memory, and behavior. They interact. They argue. They form opinions. Then you watch what happens next. This is MiroFish, and it just exploded to 36,500 GitHub stars.

March 20, 2026 · Espen · 12 min read
Traditional forecasting looks at what happened before and extrapolates forward. MiroFish does something fundamentally different.

It builds a miniature society of AI agents, gives them information, and watches how they behave. The prediction doesn't come from a statistical model — it emerges from simulated human behavior at scale. Think of it as a wind tunnel for decisions, except instead of testing aerodynamics, you're testing how people react.

What MiroFish Actually Does

MiroFish calls itself a "swarm intelligence engine, predicting anything." That sounds like marketing. But what it actually does is genuinely remarkable.

Here's the process in plain language:

  1. You provide seed material. This can be anything: news articles, financial data, policy drafts, market research, customer surveys, even novels. Plus a natural language question — "What happens to public sentiment if this policy passes?" or "How will the crypto market react to this regulation?"
  2. MiroFish creates thousands of AI agents. Each agent gets an individual personality profile, long-term memory, behavior patterns, and decision-making logic. They're not identical copies. They have different backgrounds, different opinions, different risk tolerances — modeled on the diversity of real human populations.
  3. The agents are placed in a simulated world. They interact with each other and with the seed data. They form opinions. They influence each other. They change their minds. Group dynamics emerge naturally — factions, consensus, contrarian positions, viral ideas.
  4. MiroFish observes what happens. The output is two things: a prediction report summarizing the emergent behavior and likely outcomes, and an interactive digital world you can explore — watching individual agents, tracking how opinions spread, seeing which groups formed and why.

The prediction doesn't come from a formula. It comes from emergent behavior — thousands of simulated individuals reacting to information the way real people would, and the aggregate of those reactions producing a forecast.

Why this is different from ChatGPT: If you ask ChatGPT "what happens if X?", it gives you one answer based on its training data. MiroFish creates a thousand different people and watches what they do. You don't get one prediction — you get a distribution of outcomes, weighted by how the simulated population actually behaved. It's the difference between asking an expert and running an experiment.

How the Simulation Works

Under the hood, MiroFish uses a multi-agent simulation architecture. Each agent is a lightweight LLM-powered entity with three key properties:

Individual Personality

Each agent has a unique personality profile — optimistic or pessimistic, risk-seeking or risk-averse, leader or follower, early adopter or skeptic. These aren't random assignments. MiroFish generates personality distributions that mirror real population demographics based on the context you provide. A simulation about financial markets gets agents that think like traders, retail investors, institutional managers, and regulators. A simulation about public policy gets agents that think like citizens, activists, politicians, and bureaucrats.

Long-Term Memory

Agents remember what happened earlier in the simulation. If Agent 47 read a bearish news article in round 1 and then heard Agent 200 express optimism in round 3, those memories accumulate and influence Agent 47's decisions in round 5. This is powered by Zep memory — a persistent memory layer that gives each agent a coherent history rather than treating each interaction as isolated.

Behavior Logic

Agents don't just form opinions — they act. They share information. They try to persuade others. They change their behavior based on what they see the group doing. Some agents are contrarians who push back against consensus. Some are followers who amplify popular views. The behavior logic creates realistic social dynamics where ideas spread, factions form, and consensus either builds or fractures — just like in real groups.

The simulation runs in discrete rounds. Each round, agents process new information, interact with nearby agents, update their beliefs, and take actions. MiroFish uses the OASIS simulation engine (built on CAMEL-AI) to orchestrate these interactions at scale — managing thousands of concurrent agent processes, their memory states, and their social graphs.

The result is emergent intelligence. No single agent knows the "answer." The answer emerges from the collective behavior of thousands of agents, each making individual decisions based on their unique personality, memory, and social context.

God View: Injecting Variables Into the Simulation

This is where MiroFish goes from interesting to powerful.

"God view" lets you intervene in a running simulation. You can inject new variables — a breaking news event, a policy change, a competitor's announcement — and watch the entire simulated world react in real time.

Imagine you're simulating how customers react to a price increase. The initial simulation shows moderate pushback. Now you inject a variable: a competitor drops their price by 20%. You watch the simulation evolve. Agents who were tolerating your price increase suddenly start considering alternatives. The sentiment shifts. New factions form around switching costs vs. brand loyalty.

You can run this scenario dozens of times with different variables. What if the competitor's price drop is 10% instead of 20%? What if you counter with a loyalty discount? What if you bundle a new feature? Each variable injection produces a different simulation trajectory — and the aggregate of those trajectories gives you a map of possible futures, not just one prediction.

Think of God view as A/B testing for strategy. Instead of testing two versions of a webpage, you're testing two versions of reality. Instead of measuring click rates, you're measuring how thousands of simulated stakeholders react to different scenarios. The feedback is immediate, the cost is marginal, and you can test scenarios that would be impossible or reckless to test in the real world.

Real Use Cases (From Financial Markets to Fiction)

MiroFish's versatility is part of what's driving its explosive GitHub growth. Here are documented use cases from the community:

Financial Markets

A popular fork (BTC Fear and Greed Index simulator) feeds MiroFish crypto market data and social media sentiment. Thousands of simulated traders — with different risk profiles, time horizons, and information sources — buy, sell, and hold. The emergent market behavior produces fear/greed predictions that account for crowd psychology, not just price charts. Multiple community members are tracking its accuracy against real market movements.

Public Sentiment Analysis

Feed MiroFish a proposed government policy and demographic data. Watch how different population segments react. See which groups support it, which oppose it, and — critically — see how those positions evolve as agents discuss and debate with each other. The prediction isn't just "60% approve." It's a dynamic map of how approval forms, where resistance clusters, and what arguments gain traction.

Geopolitical Scenario Planning

Specialized forks use MiroFish to simulate international relations. Agents represent nations, factions, and interest groups. Feed it a geopolitical event and watch alliance structures shift, economic responses cascade, and diplomatic positions harden or soften. Multiple research groups are using this for conflict analysis and trade policy modeling.

Fiction and Narrative Prediction

In one of the more surprising applications, MiroFish was used to predict the lost ending of "Dream of the Red Chamber" — one of China's Four Great Classical Novels. The simulation populated a world with characters from the novel, gave them their known personalities and motivations, and let the story play out. The predicted ending reportedly aligned with scholarly theories about the author's intentions. It's a fascinating demonstration of what happens when you treat narrative as emergent behavior rather than authored plot.

AI agents are transforming business operations across the board. If you're exploring how autonomous AI can grow your business, see how 10,000+ AI skills connect through MCP to give agents like Claude Code access to your entire tool stack.

How Business Owners Can Use This

MiroFish is built by researchers, but the applications for business owners are immediate and practical.

Product Launch Scenario Testing

Before you launch a new product or feature, feed MiroFish your product description, target market demographics, competitor landscape, and pricing. Watch how thousands of simulated customers react. Do they adopt it? At what price point does adoption drop? Which segments convert first? Which objections come up most often? This is focus group testing at scale — thousands of participants, instant results, and the ability to test variations without the cost and time of real research.

Pricing Strategy Simulation

Pricing is one of the highest-leverage decisions a business makes, and one of the hardest to test. MiroFish lets you simulate price changes against a population of agents modeled on your customer base. Test a 10% increase, a 20% increase, tiered pricing, freemium models — and see not just whether agents buy, but how they talk about the change, whether they churn, and how long it takes the market to stabilize.

Competitive Response Modeling

You're about to enter a new market. How will incumbents respond? Feed MiroFish the competitive landscape — existing players, their market positions, their known strategies — and announce your entry. Watch how simulated competitors react. Do they cut prices? Do they launch counter-products? Do they ignore you? The simulation surfaces responses you might not have anticipated, giving you time to prepare before you're in the market for real.

Crisis Communication Testing

Your company faces a PR crisis. You have three possible response strategies. Feed each one to MiroFish along with your customer base profile, media landscape, and social media dynamics. Watch how each response plays out in the simulated public sphere. Does the apology satisfy people or does it backfire? Does silence let the story die or does it fester? You can test crisis responses in a simulation before you test them in reality — where mistakes are permanent.

The common thread across all of these: MiroFish lets you test decisions in a simulated world before committing to them in the real one. The simulation isn't perfect — no model is — but it surfaces dynamics, reactions, and second-order effects that spreadsheet analysis and gut instinct miss.

The Tech Stack (Simplified)

For the technically curious — or for the conversation with your technical team — here's what MiroFish is built on:

The project is open source at github.com/666ghj/MiroFish with 36,500+ stars. It's backed by Shanda Group and originated from a Chinese research team, though the codebase and documentation include English, and the fork ecosystem spans Korean, German, and English-language projects.

The Fork Ecosystem

One of the best indicators that an open source project has real substance: the fork ecosystem. MiroFish has spawned a remarkable number of specialized variants:

The fork ecosystem tells you something important: developers and researchers see MiroFish as a platform, not just a tool. They're building on it because the core simulation engine is flexible enough to support radically different applications — from predicting Bitcoin prices to predicting how a novel ends.

The Honest Limitations

MiroFish is impressive. It's also not magic. Here's what you should know before you stake business decisions on it:

Simulated agents are not real people. No matter how sophisticated the personality modeling, AI agents don't have real emotions, real financial pressures, or real social relationships. They approximate human behavior based on patterns in training data. The simulation surfaces plausible dynamics, not guaranteed outcomes. Treat results as one input to your decision-making, not the decision itself.
LLM costs scale with simulation size. Running thousands of agents through multiple simulation rounds means thousands of LLM API calls. A large simulation with GPT-4 class models can cost hundreds of dollars per run. Budget models and local inference (Ollama) reduce costs dramatically but may produce less nuanced agent behavior. Plan your inference budget before running large simulations.
Garbage in, garbage out. The quality of MiroFish's predictions depends entirely on the quality of your seed data. Feed it comprehensive, balanced information and you get nuanced simulations. Feed it biased or incomplete data and the simulation will reflect those biases. This is true of every prediction tool, but it's especially important here because the emergent behavior can amplify biases in ways that are hard to detect.
Setup is non-trivial. MiroFish requires Python, Node.js, API keys, and some configuration. It's not a SaaS product you sign up for. If you don't have a technical team member who can set up a Python environment and configure API integrations, you'll need help getting started.
Early-stage project. Despite the star count, MiroFish is still evolving rapidly. APIs change, documentation lags behind the code, and some features are experimental. The community is active and helpful, but expect to troubleshoot.

Frequently Asked Questions

Q: What is MiroFish and how does it work?

MiroFish is an open source multi-agent prediction engine. You feed it seed data — news articles, policy drafts, market data, or any text — along with a natural language prediction request. It creates thousands of AI agents, each with individual personalities, long-term memory, and behavior logic, and places them in a simulated digital world. The agents interact freely, and the emergent group behavior produces predictions. MiroFish returns both a prediction report and an interactive digital world you can explore.

Q: Is MiroFish free to use?

MiroFish is open source and free to download. Running it requires LLM inference — you need an OpenAI-compatible API key (OpenAI, Anthropic, or a local model via Ollama). The cost depends on simulation size. A small run with a few hundred agents using a budget model might cost a few dollars. A large simulation with thousands of agents using frontier models could cost significantly more. Offline forks exist that run entirely on local hardware.

Q: What can MiroFish predict?

Documented use cases include public sentiment shifts, financial market behavior (there's a BTC Fear and Greed Index fork), geopolitical outcomes, and even fictional narratives — it famously predicted a lost ending of Dream of the Red Chamber. The key insight is that predictions emerge from simulated human behavior, not statistical extrapolation. Any scenario involving how groups of people respond to information is a potential use case.

Q: How can businesses use MiroFish for scenario planning?

Feed it your scenario and relevant context — a pricing change, a product launch, a competitive move — along with data about your market and customer base. Watch how thousands of simulated stakeholders react. The "God view" feature lets you inject variables dynamically — change one assumption and watch the entire simulation evolve. Use it for product launch testing, pricing strategy, competitive response modeling, and crisis communication planning.

Free: The AI Growth Breakdown

See how one business went from 0 to 100+ daily visitors in 14 days using AI agents. The exact tools and results.

Get the Free Breakdown →