MiroFish: An AI Engine that Simulates Thousands of Fake People to Predict the Future

OpenClaw gives AI agents a body. MiroFish gives them a society. One of these got the headlines, but the other might matter more.

“A Simple and Universal Swarm Intelligence Engine, Predicting Anything” — Mirofish: created by Guo Hangjiang

A 20-year-old undergraduate in Beijing vibe-coded a swarm intelligence engine in 10 days.

It hit #1 on GitHub’s global trending list. Above OpenAI, Google, and Microsoft repositories.

Within 24 hours, Chen Tianqiao, formerly the richest person in China, committed $4.1 million to incubate it.

One developer plugged it into a Polymarket trading bot, simulated 2,847 digital humans before every trade, and reported $4,266 profit over 338 trades.

The project is called MiroFish. And it does something fundamentally different from every other AI tool you’ve used.

It doesn’t ask a model what will happen. It spawns thousands of autonomous AI agents with unique personalities, memories, and social connections, drops them into a simulated world, and watches what emerges.

The emergent behavior is the prediction.

Why this is the next paradigm

Jensen Huang laid out where this is heading in his recent conversation with Lex Fridman:

[embed]~27:11–27:47

“We have now created one agentic person, and that one agentic person has a large language model that we’ve now developed. But during test time, that agentic system goes off and does research and bangs on databases, and it goes out and uses tools, and one of the most important things it does is spins off and spawns off a whole bunch of sub-agents. Which means we’re now creating large teams. It’s so much easier to scale NVIDIA by hiring more employees than it is to scale myself.

And so the next scaling law is the agentic scaling law. It’s kind of like multiplying AI. Multiplying AI, we could spin off agents as fast as you want to spin off agents.”

Huang calls this the fourth scaling law. Pre-training, post-training, and test-time compute each hit diminishing returns. The breakthrough is multiplying the number of AIs working together.

NVIDIA’s OpenClaw announcement got the spotlight. New operating system for robotics. The “iPhone of tokens.” Big, flashy, impossible to miss.

MiroFish is the quieter moment. And it might matter more.

Because while OpenClaw gives agents a physical body, MiroFish gives them a society. Thousands of agents with their own memories, opinions, and social connections, interacting across simulated platforms and producing predictions that no single model could generate alone.

How MiroFish works

MiroFish is powered by OASIS (Open Agent Social Interaction Simulations) from CAMEL-AI. It can scale to one million agents with 23 different social actions: following, commenting, reposting, liking, muting, searching.

OASIS handles the environment logic, the recommendation systems, the time engine that activates agents on schedules, and the inference layer that distributes LLM calls.

MiroFish wraps this in a clean five-stage pipeline.

Build the world. You feed it seed material: news articles, financial reports, policy drafts. MiroFish uses GraphRAG to extract entities and relationships into a knowledge graph. This becomes the shared reality all agents reason over.

Populate it. The system generates thousands of agent personas. Each gets a unique identity, stance, behavioral logic, and social connections derived from the graph. Every agent gets a persistent memory handle in Zep Cloud.

Let them interact. Simulations run on two platforms simultaneously: a Twitter-like microblogging layer and a Reddit-like forum layer. Each round, agents observe their feeds, retrieve long-term memories, invoke the LLM to decide what to do, and act. Every action is logged.

Analyze the trace. A specialized ReportAgent inspects the full simulation. It identifies what coalitions formed, where sentiment shifted, what caused inflection points, and what the plausible futures look like. The ReportAgent is architecturally separated from the agents it analyzes. Different prompts, different tool access, different governance.

Let users explore. A “God’s-eye view” lets you talk to any agent after the run, inject new variables like policy changes or macro shocks, and re-run scenarios to watch how things change.

MiroFish runs five stages in sequence. Seed documents become a knowledge graph, agents populate it, OASIS simulates their interactions, a ReportAgent analyzes the full trace, and users can re-run scenarios with new variables injected.

Emergent behavior is the prediction

This is the part that matters.

When thousands of agents with different biases interact on realistic social platforms, patterns emerge that nobody programmed. Echo chambers form. Narratives cascade through network clusters. Opinions polarize or converge depending on the social graph.

Short-form hot takes spread differently than threaded discussion. That’s why MiroFish runs both platforms at once. Cross-platform interactions create bridges between communities that would otherwise stay isolated.

Each agent carries persistent memory. Round 15 depends on what happened in Round 3. Nothing is stateless. The simulation builds on itself the same way real discourse does.

The prediction is the emergent behavior. You don’t ask the system what will happen, it runs a mass simulation of language models that interact. You watch it happen.

Agents cluster into coalitions on two parallel platforms. Bullish and bearish groups form organically through interaction, while cross-platform links (purple) create information bridges between communities that would otherwise stay isolated.

The architecture worth stealing

For builders, the real value here is the modular decomposition. Every piece is swappable.

The world model is a knowledge graph built by GraphRAG. Swap it for any structured data store. The agent generator creates personas from graph data with clear schemas. OASIS handles simulation mechanics. Zep Cloud handles persistent memory. The ReportAgent sits in its own layer with its own tooling.

This separation is critical. When you tangle participants and analyzers in the same system, outputs get muddled. MiroFish keeps them clean.

The system is model-agnostic. It defaults to Qwen-plus via Alibaba’s Bailian platform using an OpenAI-compatible API schema. Any sufficiently capable LLM can serve as the agent brain. Swap in GPT, Claude, or a local model without touching the architecture.

The modular architecture separates simulation core from external services. Teal components handle world-building and agent generation. Coral components handle analysis. Every piece is swappable: replace the LLM, the memory service, or the simulation engine without touching the rest.

The tradeoffs

Powerful, but expensive. Hundreds of agents over dozens of rounds means tens of thousands of LLM calls. That can translate to hundreds or thousands of dollars per simulation without careful constraints.

Cost optimization strategies include model routing (cheaper models for background agents, premium for the ReportAgent), prompt caching, and using GraphRAG to reduce context bloat.

There’s also a calibration issue. LLM-based agents tend to display exaggerated herd behavior compared to real humans. Simulations can overestimate how fast sentiment swings. The MiroFish team is upfront about this: outputs are synthetic scenarios for structured thinking, not ground truth.

Calibration against historical events and triangulation with other methods remain essential. But the Polymarket results suggest the signal-to-noise ratio is already actionable.

Why you should pay attention

The single-model era of AI is ending.

Huang is calling agentic scaling the fourth scaling law. NVIDIA is building hardware specifically designed for it. And a 20-year-old just proved you can build the whole thing in 10 days with open-source tools and standard LLM APIs.

That gap between the vision and the accessibility is where the opportunity lives.

You could build a version of this for trading sentiment with 500 agents representing different market participants. Product launch scenarios with agents mapped to customer segments. Political forecasting with agents tied to demographic clusters. Reputation analysis. Crisis communication planning. Internal alignment exercises.

The patterns transfer. The infrastructure is commodity. The hard part is knowing what questions to ask, how to validate the outputs, and how to act on what the swarm reveals.

The information is there if you know where to look.

MiroFish is one more tool for looking.