MiroFish: when thousands of AI agents simulate the world to predict what happens next

A project built in 10 days that hit #1 on GitHub global trending

In March 2026, a 20-year-old from Beijing named Guo Hangjiang published a project on GitHub called MiroFish. He had built it in ten days using a technique now well known in the developer community: vibe coding — generating code with AI and iterating rapidly. Within days, the repository reached 43,000 stars and 5,900 forks, landing at #1 on GitHub Global Trending, above repositories from OpenAI, Google, and Microsoft.

Chen Tianqiao — formerly China's richest person and founder of the Shanda Group — committed $4.1 million to incubate it within 24 hours of the first rough demo going live. Guo went from university student to CEO overnight.

It is worth understanding what this project actually does, because it touches on some of the most interesting ideas in the current AI landscape.

What are Agent Swarms

Before MiroFish, you need to understand the concept it is built on: the agent swarm.

A single AI agent — an advanced chatbot, an autonomous assistant — receives a task, reasons, uses tools, produces an output. It works well for linear tasks. But many real-world phenomena are not linear: public opinion, financial markets, social dynamics are all the result of millions of interactions between individuals who influence each other.

An agent swarm is a system where not one agent runs, but thousands or millions simultaneously, each with its own personality, its own memory, and its own behavioral logic. They are not orchestrated by a central controller to complete a step-by-step task. They are released into a simulated environment and interact freely with one another.

The result is emergent behavior: collective patterns that were not explicitly programmed into any single agent, but that arise from interaction. Exactly as happens in real human societies.

This is a paradigm shift compared to the AI most developers are familiar with.

What MiroFish does

MiroFish's premise is this: to predict what will happen in a complex system — a market, a public opinion landscape, the ending of a novel — you do not ask a language model to reason about the problem. You build a digital copy of that system, populate it with thousands of agents representing its actors, and watch them interact.

The workflow is structured in five phases:

1. Graph Building. The source material is loaded — a report, a news article, a literary text. The system processes it through GraphRAG, which extracts entities and relationships to build a knowledge graph of the domain.

2. Environment Setup. Agents are generated from the graph: each with a unique personality, a social history, pre-existing relationships with other agents, and specific behavioral parameters.

3. Simulation. Agents are released into an environment that simulates social platforms — a Twitter-like environment and a forum-like environment — where they post, comment, follow, debate, and shift their positions over time. Each agent's temporal memory is updated dynamically throughout the simulation. The underlying engine is OASIS, an open-source framework by the CAMEL-AI team for multi-agent social simulations.

4. Report Generation. A ReportAgent with access to the entire post-simulation environment produces a structured report: not just where sentiment ended up, but how the different factions formed, shifted, and influenced each other along the way.

5. Deep Interaction. You can converse with any agent in the simulation, or ask further questions to the ReportAgent.

Agents' long-term memory is managed through Zep Cloud. The technical stack is Python for the backend, Vue for the frontend, with Docker support and source deployment. The base LLM is compatible with any API in OpenAI SDK format.

Documented experiments

This is not just architecture on paper. In the weeks following publication, a number of concrete experiments have already emerged.

Public opinion forecasting. The team's first official demo simulates sentiment dynamics around a controversy at Wuhan University. MiroFish, fed with a public opinion analysis report generated by BettaFish (Guo's previous project), simulated how sentiment would evolve over the following weeks, showing the formation and shifting of different factions.

Literary narrative prediction. The second demo is conceptually bolder: the team loaded the 80 chapters of Dream of the Red Chamber — the great 18th-century Chinese classic novel, left without a finished ending — and asked MiroFish to simulate how the story might have concluded. Thousands of agents, each representing a character with personality and relationships drawn from the text, played out the remaining narrative arc.

Prediction market trading. One developer connected MiroFish to a Polymarket trading bot: it simulated 2,847 digital humans before each trade and reported a profit of $4,266 across 338 trades.

Scale test. Brian Roemmele successfully ran a single MiroFish simulation with 500,000 AI agents.

Possible use cases

The use cases documented or declared by the team cover a wide range:

• Communications and PR: simulate how the public will react to an announcement, a crisis, or a campaign before publishing it. Zero real-world risk, data emerging from simulation.
• Policy testing: test the perceived impact of a proposed law or strategic decision across different segments of a simulated population.
• Financial markets: not deterministic price prediction, but simulation of group sentiment under different scenarios — probabilistic thinking applied to trading.
• Academic research: simulate social dynamics, information diffusion, and political polarization on controlled synthetic populations.
• Narrative and game design: explore alternative story arcs, test character consistency, simulate player reactions.

A fork already exists — MiroFish-Offline — which replaces Zep Cloud with a local Neo4j instance and uses Ollama, allowing simulations to run without any external API.

Why it is interesting

MiroFish addresses a problem that classical language models struggle with: complex adaptive systems. An LLM reasoning about "how will public opinion around X evolve" produces an analysis, but that analysis is fundamentally its own projection based on text patterns. MiroFish instead builds an environment, populates it with autonomous agents, and observes what emerges.

The emergent behavior is real — it is not simulated by a single model pretending to be a crowd. It is the result of thousands of agents with independent state influencing one another.

There are, however, two limitations worth acknowledging honestly.

The first is empirical validation: MiroFish is at version 0.1.2, released on March 7, 2026. No published benchmarks yet exist that compare the system's predictions against real historical outcomes at scale. The plausibility of an output is not proof of its accuracy.

The second is the dual-use question: the same engine that simulates public opinion to predict it can be used to study it for the purpose of manipulating it. The fact that the project is open-source makes this dynamic particularly open, and the team has not yet produced a systematic reflection on the point.

In summary

MiroFish is not an operational tool ready for production. It is an experimental framework with a solid architectural idea — swarm intelligence applied to the prediction of complex systems — built in record time by a young team with serious backing behind it.

It is worth monitoring if you work in contexts where collective dynamics matter: communications, finance, policy, social research. The problem it is trying to solve is real. If the emergent simulations manage to translate into verifiable predictive accuracy, this will be a category of tools with very few precedents.

For now, it is the most interesting project to emerge from the open-source AI ecosystem in the first quarter of 2026.