Mirofish is an open-source AI tool built to explore how collective human behavior might unfold under different conditions. Instead of forecasting a single numeric outcome, Mirofish runs large-scale social simulations by generating thousands of autonomous AI agents with distinct personalities, memories, and behavioral rules. These agents interact inside digital, social-media-like environments, producing emergent patterns such as opinion shifts, polarization, and coalition formation.

Launched in early March 2026 by a 20-year-old developer reportedly built in about ten days, Mirofish gained rapid traction in the developer community, reaching over 33,000 GitHub stars, topping GitHub global trending, and attracting 30 million yuan (approximately $4.1 million) in funding within 24 hours. The project builds on a predecessor called BettaFish (late 2024) and is designed to test narratives and strategies by simulating many plausible futures rather than claiming certainty about any single outcome.

What is Mirofish?

Mirofish is described as an AI swarm intelligence prediction engine. The core idea is straightforward: complex social outcomes often emerge from many individual interactions. Traditional models frequently struggle to capture herd effects and feedback loops because they treat populations as aggregates. Mirofish attempts to model those dynamics directly by creating a population of agents and letting them interact in a controlled environment.

Many observers compare the concept to a "SimCity for prediction" because users can run scenarios, change variables, and observe how the simulated society evolves. This makes Mirofish relevant for professionals who need to stress-test decisions where human reactions matter, including marketing, investor communications, crisis management, and policy analysis.

How Mirofish Works: The Five-Stage Pipeline

Mirofish runs a multi-step workflow that transforms documents into a simulated social world and then into structured outputs. At a high level, it focuses on extracting knowledge, instantiating agents, simulating interactions at scale, and summarizing emergent patterns.

1) Knowledge graph construction with GraphRAG

Mirofish begins by ingesting input documents such as news articles, financial reports, policy drafts, or novels. It then constructs a knowledge graph using GraphRAG, extracting entities and relationships. This step provides the world model that agents reference when forming opinions, debating, or coordinating.

2) Agent creation with personalities, memory, and behavior rules

Mirofish generates autonomous agents with distinct profiles and behavioral logic. A key feature is long-term memory supported through Zep Cloud, which helps agents remain consistent over time rather than behaving like stateless chatbots.

In practice, this stage creates:

Diverse personas (for example, skeptical analysts, enthusiastic fans, cautious regulators)
Persistent memory so agents can reference prior interactions and update beliefs
Behavioral tendencies that shape posting frequency, conformity, contrarianism, and influence

3) Social simulation on parallel platforms via OASIS (CAMEL-AI)

Once agents exist, Mirofish simulates interaction across parallel social environments that resemble Twitter-like and Reddit-like platforms. This simulation layer is powered by the OASIS engine from CAMEL-AI and is notable for its scale. Reported capabilities include support for up to one million agents and 23 types of social actions such as posting, commenting, and following.

This is where emergent behavior appears. Agents influence each other, propagate narratives, form groups, and react to events. The goal is not to produce a single prediction, but to reveal possible trajectories and tipping points.

4) ReportAgent: structured analysis of emergent patterns

After the simulation runs, a ReportAgent summarizes what happened and converts it into structured predictions. Rather than producing a raw transcript of agent activity, this stage identifies patterns that decision-makers care about, including:

Opinion drift and sentiment direction over time
Coalition formation and the emergence of distinct camps
Polarization and fragmentation dynamics
Influencer effects and information cascades

5) Deep interaction and scenario replays

Mirofish is designed for iterative exploration. Users can query agents, inspect reasoning, or inject new variables to rerun simulations. This "what if we change X?" workflow is central to how the tool is positioned: it is more aligned with narrative stress-testing than with point forecasting.

Tech Stack and Licensing

From a practical adoption standpoint, Mirofish is built with:

Python 3.11+ for the backend
Vue.js for the frontend
AGPL-3.0 open-source license

The AGPL-3.0 license carries implications for enterprises because it can impose source-sharing obligations when the software is used over a network. Teams evaluating Mirofish should involve legal and compliance stakeholders early, particularly if they plan to modify and deploy it internally or as part of a commercial service.

Where Mirofish Fits: Use Cases That Benefit From Swarm Simulation

Mirofish is most useful when outcomes depend on how groups respond, coordinate, or polarize. The following examples illustrate where a swarm-based AI tool can provide insights that static surveys or simple time-series models may miss.

Finance and market narrative stress-tests

Mirofish has been discussed in the context of modeling how different market participants might react to events such as a Federal Reserve rate decision. The goal is not to predict the exact price of an asset. Instead, it can help teams explore how sentiment might converge or diverge across:

Retail investors
Institutional investors
Analysts and commentators

This can support scenario planning, communications strategy, and monitoring potential feedback loops in market narratives.

Marketing and brand strategy

Marketing teams typically rely on surveys, focus groups, and A/B tests. Mirofish-style simulations offer a different perspective: observe how messaging propagates through a network with heterogeneous personas and social influence dynamics. This can help teams identify:

Which segments amplify or resist a message
How backlash narratives could emerge
What types of content become persistent inside specific communities

Public opinion and crisis dynamics

For public opinion analysis, Mirofish can simulate sentiment evolution following a news event, including how controversy spreads and which groups form around competing interpretations. For crisis response teams, this can support tabletop exercises by helping planners identify likely failure modes and understand second-order effects.

Creative and cultural prediction

One publicly discussed example involves applying Mirofish to the first 80 chapters of Dream of the Red Chamber to simulate a plausible lost ending. While creative use cases are less enterprise-focused, they demonstrate that the engine can be applied to narrative continuation and cultural dynamics beyond corporate scenarios.

Benefits and Limitations: How to Use Mirofish Responsibly

Mirofish's value lies in modeling interaction effects: cascades, bandwagoning, echo chambers, and coalition building. At the same time, users should treat outputs as hypothetical and avoid overconfidence in any single simulation result.

Key strengths

Captures emergent behavior that is difficult to observe in aggregated models
Scales substantially with reported support for up to one million agents
Iterative scenario exploration through replays and variable injection
Open-source transparency for inspection, adaptation, and peer review

Key risks and gaps to watch

Lack of published benchmarks and limited standard accuracy metrics for real-world validation
Speculative narratives can be misinterpreted as predictions with certainty
Security and governance concerns when deploying agent-based systems inside enterprises
License implications under AGPL-3.0 for networked deployments

For enterprise users, a practical approach is to treat Mirofish as a decision-support simulator rather than an oracle. Pair it with human review, traditional analytics, and controlled evaluation methods such as backtesting on historical scenarios where feasible.

How AI Users Can Get Started With Mirofish

For researchers, product strategists, and analysts exploring Mirofish, disciplined experimentation produces the most reliable insights:

Start with a narrow question, such as "How might sentiment evolve if we announce X?"
Curate inputs carefully (news articles, reports, briefings) because the knowledge graph directly influences agent beliefs
Run multiple simulations with different seeds and parameters to compare trajectories
Track outputs systematically by noting common patterns, outliers, triggers, and influencer nodes
Document assumptions about agent personas, incentives, and platform rules

Building a strong foundation in applied AI helps practitioners evaluate tools like Mirofish more effectively. Blockchain Council offers relevant training paths including the Certified Artificial Intelligence (AI) Expert for applied AI foundations, Certified Machine Learning Expert for modeling rigor, and Certified Data Science Professional for evaluation frameworks and responsible interpretation.

Conclusion: Why Mirofish Matters for Scenario Planning

Mirofish represents a notable shift in how AI users can approach prediction: away from single-number forecasts and toward simulation of collective behavior. Its pipeline combines knowledge graphs, long-term agent memory, and large-scale social interaction modeling to generate structured narratives about how groups might react under different conditions.

Its outputs should be treated as exploratory rather than definitive. Without standardized validation results and published benchmarks, the most responsible use is as a scenario laboratory for stress-testing strategies, anticipating second-order effects, and improving preparedness. For practitioners who understand both the capabilities and the limitations, Mirofish offers a structured way to think in systems rather than in spreadsheets.