One Undergrad Built 700,000 AI Agents in 10 Days That Predict Markets and Elections

An undergraduate student built MiroFish in 10 days using vibe coding. It uses swarms of up to 700,000 AI agents -- each with its own memory, personality, and behavioral logic -- to simulate how crowds think and predict what they will do next. The live demos were described as "scarily accurate" by tech commentator Muhammad Ayan. Within days of launch, MiroFish topped GitHub's global trending list with over 16,000 stars. Shanda Group founder Chen Tianqiao invested 30 million RMB.

This is not a prediction market. It is not a polling tool. It is a parallel digital world where AI agents live, argue, react to news, and influence each other -- and the patterns that emerge mirror what real crowds do before they do it.

What Is MiroFish?

MiroFish is an open-source swarm intelligence engine that simulates collective human behavior using thousands of autonomous AI agents to predict market movements, public opinion shifts, and narrative outcomes. It was built by Guo Hangjiang, an undergraduate developer, and published on GitHub at 666ghj/MiroFish. The project is 100% free and open source.

Each agent in the swarm is not a simple rule-following bot. Every agent carries independent personality traits, long-term memory of past events, and behavioral logic that evolves as the simulation runs. Agents communicate with each other, react to injected information, and influence each other's decisions -- producing emergent patterns that the creator claims mirror real crowd dynamics before they manifest in the real world.

How MiroFish Works: The Technical Architecture

MiroFish runs a four-stage pipeline to turn real-world data into crowd simulations:

Stage 1 -- Knowledge extraction: GraphRAG (Graph Retrieval-Augmented Generation) processes seed data -- breaking news, policy drafts, financial signals, social media feeds -- and extracts entities, relationships, and context into a structured knowledge graph. This is what the agents "know" about the world when they wake up.

Stage 2 -- World construction: From the knowledge graph, MiroFish automatically builds a high-fidelity digital environment. If you feed it information about a proposed central bank rate decision, it constructs a simulated financial world populated with agents who represent different market participants -- retail investors, institutional traders, analysts, journalists.

Stage 3 -- Agent simulation: The engine spins up between thousands and 700,000 agents (up to 1 million in some configurations). Each agent has a distinct personality profile, memory of prior interactions, and behavioral tendencies. Some agents are risk-averse. Some are contrarian. Some follow crowd sentiment. Some act on incomplete information. The diversity is the point -- it mirrors real market heterogeneity.

Stage 4 -- Dynamic variable injection: Users can inject new variables from a "God's-eye view" -- a policy announcement, a corporate earnings miss, a geopolitical event -- and watch the simulation respond. The agents react, communicate, shift behavior, and produce measurable aggregate outcomes. The output is a probability distribution of future states, not a single prediction.

What MiroFish Can Predict

The system is designed for macro-level crowd behavior prediction rather than individual outcomes:

Muhammad Ayan, who observed the live demos, described the accuracy as "scarily accurate" -- specifically noting its ability to simulate how crowds think before it happens. Brian Roemmele, founder of The Zero-Human Company, who began implementing MiroFish immediately, said: "We used a multi-agent AI engine to simulate parallel digital worlds with 700,000 diverse agents to predict outcomes like market trends or public reactions for emergent social behaviors. This is a massive superpower."

The Creator: An Undergrad, 10 Days, Vibe Coding

Guo Hangjiang built MiroFish in 10 days. He is an undergraduate student. The development method was vibe coding -- using AI coding assistants to accelerate implementation rather than writing every line manually.

This detail matters for two reasons. First, it demonstrates how much the ceiling for individual developer ambition has shifted in 2026. A system of this architectural complexity -- GraphRAG integration, multi-agent orchestration, dynamic world simulation -- would have required a funded team and months of engineering a few years ago. Second, it signals that MiroFish is not a polished commercial product. It is a proof of concept built by one person in under two weeks, which explains both its rapid GitHub traction and its current limitations.

The 30 million RMB investment from Chen Tianqiao (founder of Shanda Group, one of China's pioneering gaming and internet companies) suggests the technology has commercial development ahead of it. But as of March 2026, what is on GitHub is a raw, powerful, and rough engine.

How Developers Can Use MiroFish Right Now

MiroFish is open source on GitHub. Getting started requires:

• Python environment (the core engine is Python-based)
• An LLM API key (OpenAI, Claude, or compatible endpoint) -- agents are powered by language models
• Seed data for your simulation domain (news articles, financial data, policy documents)

The basic workflow:

1. Prepare seed data describing the scenario you want to simulate
2. Configure agent count and personality distribution profiles
3. Run the world construction pipeline to build the digital environment
4. Inject your scenario variable (the event you want to test)
5. Run the simulation and collect aggregate behavioral output

For developers building in prediction markets, financial analytics, political research, or social media analysis, MiroFish offers a simulation layer that no API currently provides. Instead of asking "what do people think," you can ask "what will people do when X happens" and simulate the answer at scale.

The current limitations to be aware of: the system is compute-intensive at scale (700,000 agents requires significant GPU/CPU resources), the documentation reflects 10-day development pace, and LLM API costs accumulate fast when running large swarms.

What MiroFish Cannot Do: The Honest Limitations

The "scarily accurate" framing deserves scrutiny.

MiroFish will not reliably predict short-term price movements. The creator explicitly states this -- asking it whether BTC will go up or down tomorrow is outside its design intent. It models crowd behavior patterns, not deterministic outcomes.

The deeper limitation is fundamental to any agent-based simulation: agents are not people. Their personality settings are derived from LLM training data, which carries systemic biases. Real human behavior has chaotic elements -- personal circumstances, off-model reactions, genuine novelty -- that no training dataset fully captures. The simulation is a map, not the territory.

There is also an adversarial use concern. A tool that accurately models how narratives spread through crowds is, by design, also a tool for testing how to most effectively inject those narratives. The same engine that predicts public opinion can be used to engineer it. The project being open-source makes this dual-use dynamic particularly sharp.

The swarm intelligence market is projected to reach $7.23 billion by 2032 at a 41.2% CAGR. MiroFish is the first open-source implementation to demonstrate this capability at scale. Whether the current accuracy holds under rigorous benchmarking -- against actual market outcomes, actual election results -- is not yet established. The demos are impressive. The independent validation is not yet there.

Why This Changes the Developer Landscape

MiroFish represents a shift in what individual developers can build. The architecture -- GraphRAG for knowledge extraction, autonomous agents with memory, emergent interaction patterns -- has been described in research papers for years. Building a working implementation at this scale, in 10 days, by one person, was not previously achievable.

For developers building in 2026: the playbook for AI applications is expanding. You no longer need a large dataset of historical outcomes to build a prediction tool. You can simulate the crowd, inject the variable, and observe the emergent behavior. The question of whether those simulations are accurate enough to be useful is going to be answered over the next 12 months as teams build on top of this foundation.

Guo Hangjiang uploaded an engine. What developers build on top of it is the next chapter.

Key Takeaways

• MiroFish is an open-source swarm AI engine built in 10 days by an undergraduate developer, now with 16,000+ GitHub stars and 30 million RMB in investment
• It spins up up to 700,000 AI agents -- each with individual memory, personality, and behavioral logic -- to simulate crowd behavior and predict outcomes
• Described as "scarily accurate" by tech commentator Muhammad Ayan after live demos; Brian Roemmele called it "a massive superpower" and began implementing it immediately
• Use cases: market sentiment, public opinion shifts, narrative spread, election dynamics, product launch simulation
• MiroFish will not predict short-term price movements -- it models macro crowd behavior patterns, not deterministic individual outcomes
• For developers: the project is open source on GitHub (666ghj/MiroFish); requires Python and an LLM API key; compute costs scale with agent count
• What to watch: independent accuracy benchmarking against real-world outcomes, and the dual-use implications of a tool that can both predict and engineer public opinion