MiroFish 1M Agents: Why It Fails at Trading But Wins Everything Else

MiroFish has scaled from 700,000 agents to 1,000,000 — and one developer decided to find out what one million synthetic agents actually predict correctly by running 338 live Polymarket trades. The result: $4,266 in profit before the framework hit limits that no amount of scale can fix.

The original MiroFish went viral when a Chinese research team built a swarm engine that simulated mass human behaviour at scale nobody had attempted before. The follow-up story is more interesting: the MiroF918 swarm AI launch post covered the GraphRAG architecture and why 700K agents changed social simulation. Now at 1M agents, the real-world experiments are in — and they draw a clear boundary between where swarm AI excels and where it breaks.

What Changed in the 1M Agent Version

Version 0.1.2, released in March 2026, pushed the agent ceiling from 700,000 to one million. The architecture additions:

GraphRAG memory persistence got a significant update — agents now maintain relationship maps across simulation runs rather than resetting between sessions. This matters for longitudinal social simulations where agent influence networks accumulate over time.

OASIS framework integration (from CAMEL-AI) was tightened. OASIS is the open-source social simulation backbone that MiroFish runs on top of — it handles agent scheduling, message passing, and the Twitter/Reddit-style platform environments that agents inhabit. The 0.1.2 version improves multi-platform simulation so agents can operate across simulated social environments simultaneously.

MiroFish-Offline fork: Separately, an English-language fork emerged for developers who cannot or will not use cloud APIs. MiroFish-Offline runs entirely locally — Neo4j for the knowledge graph, Ollama for inference (qwen2.5:32b is the recommended model), Docker Compose for the full stack. Zero OpenAI. Zero cloud dependencies. The setup is about 40 minutes on a machine with a capable GPU.

The Shanda Group investment came fast: Chen Tianqiao, founder of the Chinese gaming company that became an investment firm, committed $4.1M within 24 hours of the original MiroFish going viral. That kind of speed from a serious institutional investor signals that someone with resources to do due diligence saw the architecture and moved immediately.

The Polymarket Experiment: 338 Trades, $4,266 Profit

A developer documented running MiroFish against Polymarket prediction markets. The setup: feed current event context into the swarm, collect agent votes on binary outcomes, bet accordingly on Polymarket when swarm confidence exceeded a threshold.

338 trades. $4,266 in realised profit. Positive expected value — until it wasn't.

The cases where MiroFish outperformed Polymarket market prices:

Long-horizon political questions: What will US policy be in 90 days? Will a specific treaty pass? Questions where the answer depends on aggregate public opinion and institutional momentum — the kind of dynamics that a million simulated agents with realistic social network structures can model.

Corporate sentiment events: Will a company's CEO face backlash after a public statement? Will a product launch generate negative press coverage? MiroFish's strength is predicting how real humans react to information, and this is exactly that class of problem.

Slow-moving geopolitical questions: Will sanctions be extended? Will a country exit a trade agreement? These play to the same strengths — aggregate institutional and public sentiment over weeks.

The cases where it failed:

Sub-15-minute resolution markets: Polymarket offers markets that resolve in minutes based on real-time price feeds — crypto prices, specific sports scores at halftime, that kind of thing. MiroFish cannot model financial market microstructure, high-frequency order flow, or the reflexive dynamics where the prediction market itself moves the underlying. The agents are humans, not algorithmic traders.

Sports injury markets: "Will [player] start the next game?" This class of market resolves on private information — a team physician's assessment, a coach's internal decision. No amount of public sentiment simulation predicts private information.

Thin market edge cases: In very low-liquidity Polymarket markets, MiroFish's confidence thresholds picked up noise as signal. The framework needs sufficient trading volume on the other side to provide calibration.

The developer's conclusion: MiroFish has positive edge on political and social outcome markets. It has no edge on financial microstructure or private-information markets.

Why Swarm AI Cannot Beat Prediction Markets at Everything

The theoretical reason MiroFish hits a ceiling is worth understanding, because it shapes every use case decision you make with the framework.

Prediction markets aggregate the private information of real participants who have money at stake. When a sports bettor knows an injury is real because they saw the player limping at practice, that private signal gets incorporated into the market price. When a political insider knows a vote will flip because they were in the room, that signal appears in the market.

MiroFish agents have no private information. They have public information, processed through a realistic social network topology, with memory and influence dynamics that approximate how public opinion forms. That is genuinely powerful for modelling aggregate public reactions. It is useless for capturing private information that never enters public discourse.

The Polymarket results confirm this. Markets where MiroFish had edge were markets where outcomes depended on publicly-knowable social and political dynamics. Markets where it had no edge were markets where outcomes depended on information that was genuinely private until resolution.

This is not a limitation of MiroFish specifically — it is a fundamental constraint on any approach that works only from public information. It says something important about where to deploy the framework.

MiroFish-Offline: The Self-Hosted Developer Guide

The offline fork is the most practically interesting development for developers who want to run experiments without cloud costs.

Requirements:

• A machine with at least 32GB RAM (64GB recommended for comfortable operation at scale)
• GPU with 24GB+ VRAM for qwen2.5:32b at full quality (or quantised versions on 16GB)
• Docker and Docker Compose
• Neo4j Desktop or the Docker Neo4j image

Stack:

• Neo4j: The knowledge graph that stores agent memory, relationship maps, and entity context between simulation runs. Neo4j's Cypher query language is how the agents traverse their relationship networks.
• Ollama: Local LLM inference server. qwen2.5:32b is the recommended model for agent reasoning quality — it handles the complex social reasoning tasks that make MiroFish agents feel like distinct entities rather than clones.
• Docker Compose: Orchestrates Neo4j, Ollama, and the MiroFish application layer together. The setup file in the repo handles port configuration and volume mounting.

Agent count on local hardware: The cloud version reaches 1M agents because cloud infrastructure handles the parallelisation. On local hardware with qwen2.5:32b, realistic agent counts are in the 10,000-50,000 range depending on hardware. That is still large enough for meaningful social simulation — the original research showing emergent social dynamics used far fewer agents.

The privacy argument for self-hosting: if you are running competitive intelligence simulations, political research, or proprietary scenario planning, sending your context through a cloud API creates exposure. MiroFish-Offline keeps everything on your infrastructure.

What MiroFish Actually Wins At

The trading experiment scopes the limits. But the wins are real and significant.

Public opinion simulation before campaigns: Marketing and political teams can run MiroFish simulations on proposed messaging before spending. The agents predict how real social networks would respond, including negative sentiment cascade dynamics that kill campaigns.

Crisis communication modelling: What happens if this statement gets released? MiroFish can simulate the social propagation across a million agents and return the predicted sentiment curve, identifying where the backlash concentrates.

Product launch scenario planning: Will this feature get adopted or resisted? The agents have consumer psychology baked into their social network dynamics. A tech company can simulate the dev community's reaction to an API pricing change before announcing it.

Regulatory impact modelling: How will the public respond to this policy? Government agencies and advocacy groups can run MiroFish simulations of proposed regulations and model the political pressure response. The long-horizon accuracy that makes it work on Polymarket political questions is directly applicable here.

Social research at scale: The academic use case is not going away. Simulating how misinformation spreads through a social network, how activist movements gain momentum, how consumer boycotts organise — all of these are research problems that MiroFish-scale simulations can address in hours rather than years of longitudinal data collection.

The $4.1M Investment Signal

Chen Tianqiao's Shanda Group investing $4.1M within 24 hours of virality is notable not just for the speed but for the source. Shanda is not an AI lab — it is an investor with a long track record of identifying platform-layer technologies early. The bet is not on MiroFish's current capabilities; it is on the infrastructure position that a leading swarm AI engine occupies as the use cases above become standard business tools.

The investment gives the MiroFish team resources to build what the open-source version cannot: enterprise reliability, API access controls, compliance tooling for sensitive simulations, and the infrastructure to reliably serve 1M agent runs at commercial scale.

For developers, this means the open-source version is likely to remain genuinely open (the investment does not require a proprietary pivot), while a commercial tier emerges around the use cases that enterprise customers will pay for. That is the standard open-core trajectory.

Running Your Own Experiment

If you want to reproduce the Polymarket experiment or build on the offline version:

The MiroFish GitHub repository has the 0.1.2 release with full Docker Compose configuration. The offline fork documentation covers the Neo4j schema and the Ollama model configuration. For prediction market experiments, Polymarket has a public API that returns market prices and resolution data — the experiment pipeline is straightforward: pull market context, run simulation, compare agent consensus to market price, log trades.

The useful calibration question to answer first: what class of problem are you modelling? If it is a question where the answer depends on how real humans process public information and form opinions, MiroFish has genuine edge. If it is a question where the answer depends on private information or financial microstructure, no amount of agent scale helps.

Key Takeaways

• MiroFish 0.1.2 scales to 1M agents — up from 700K; GraphRAG memory persistence improved, OASIS integration tightened
• Polymarket experiment: 338 trades, $4,266 profit on political and social outcome markets — positive edge on publicly-knowable dynamics, zero edge on sub-15-minute and private-information markets
• Why markets are different: prediction markets aggregate private information that MiroFish agents do not have — the ceiling is fundamental, not a scale problem
• MiroFish-Offline fork: Neo4j + Ollama + Docker Compose, zero cloud, qwen2.5:32b recommended; 10K-50K agents on local hardware depending on GPU
• Real wins: pre-campaign opinion simulation, crisis communication modelling, product launch scenario planning, regulatory impact modelling — all domains where aggregate public reaction determines the outcome
• $4.1M from Shanda Group (Chen Tianqiao) within 24 hours of virality — open-core commercial trajectory expected
• For developers: the framework is free, the experiments are cheap on local hardware, and the class of problems it solves (social dynamics at scale) has no good alternative