Artificial intelligence is evolving fast, but most tools still operate the same way: you give a model a prompt, and it returns a response. That’s useful, but it’s limited. What if you could simulate how groups of AI agents interact, debate, and influence each other inside a digital world?

That’s the idea behind MiroFish, a multi-agent AI engine that can predict reactions to news, market shifts, policy changes, or even storylines in a novel. Instead of a single answer, MiroFish creates a dynamic, interactive society of thousands of AI agents, each with their own memory, behavior, and perspective.

💡 Pro Tip: Building or interacting with AI agents and MCP servers? Apidog provides a powerful, built-in MCP Client specifically designed for debugging and testing MCP Servers. Whether you're connecting via STDIO for local processes or HTTP for remote servers, Apidog offers an intuitive visual interface to effortlessly test executable Tools, predefined Prompts, and server Resources. It automatically handles complex OAuth 2.0 authentications and dynamically renders rich Markdown and image responses making it the ultimate tool for seamless MCP integration testing.

Unlike traditional AI tools that generate answers directly, MiroFish builds an entire digital society of AI agents. Each agent has its own memory, personality traits, and decision-making logic. When a new event is introduced such as breaking news, a policy proposal, or a financial signal the agents begin interacting with one another, reacting to the information and influencing each other’s behavior.

Over time, their interactions create patterns that resemble how real groups of people react to events. These patterns can reveal possible outcomes, emerging narratives, or shifts in sentiment, making the system a powerful environment for experimentation and forecasting.

Source: X

What Is MiroFish?

At its core, MiroFish is a swarm intelligence simulation engine built around multi-agent artificial intelligence.

Instead of relying on a single AI model, the platform generates a large population of autonomous agents that exist inside a simulated digital environment. Each of these agents represents an individual participant in a virtual society.

Every agent has its own:

• personality traits


• behavioral rules


• long-term memory


• social relationships


• decision-making processes

When agents interact with one another, they exchange information, form opinions, and respond to events. This creates emergent behavior, meaning large-scale outcomes arise naturally from many individual interactions.

The concept mirrors real human societies. In the real world, public opinion, market movements, and social trends often emerge from millions of individual decisions. By simulating these interactions digitally, MiroFish attempts to model how events may unfold before they happen.

In simple terms, the platform acts as a digital sandbox for exploring “what-if” scenarios.

The Vision: A Mirror of Collective Intelligence

The vision behind MiroFish is to create what the developers describe as a collective intelligence mirror of the real world.

Traditional predictive systems often rely heavily on historical data and statistical models. While these approaches can work well in stable environments, they often struggle when human behavior becomes unpredictable.

Many real-world events are shaped by social interactions rather than numerical patterns alone.

For example:

• financial markets can swing due to investor sentiment


• social media trends can spread unpredictably


• public reactions to policies can change rapidly

MiroFish approaches prediction differently. Instead of trying to compute the future directly from data, the system recreates a digital environment where individuals interact and influence each other.

The idea is that complex outcomes can emerge naturally from these interactions.

By observing how simulated agents respond to events, the platform can generate insights into potential real-world outcomes.

From Seed Data to a Digital World

Running a simulation in MiroFish begins with what the system calls seed material.

Seed material is the information that defines the scenario to be simulated. This could include:

• breaking news articles


• financial reports


• policy documents


• research papers


• social media discussions


• or even fictional stories

Users upload the material and describe their prediction goal using natural language.

For example, someone might ask the system to simulate:

• how markets will react to a new policy announcement


• how the public will respond to a controversial statement


• how a story might unfold if missing chapters were completed

Using this information, MiroFish constructs a digital environment where agents can begin interacting.

The system essentially creates a parallel digital world where the scenario can play out.

MiroFish Workflow: How the Simulation Pipeline Works

Behind the scenes, MiroFish follows a structured pipeline that transforms real-world data into a dynamic simulation environment. Each stage prepares the information needed for agents to interact and produce meaningful outcomes.

1. Knowledge Graph Construction

The first stage extracts seed information from real-world data sources.

These sources may include:

• breaking news events


• financial reports


• policy drafts


• research documents


• social discussions

The system then builds a knowledge graph using a GraphRAG architecture. This graph organizes entities, relationships, and contextual information that agents will use during the simulation.

In addition to structured data, both individual and group memory structures are injected into the simulation so agents can retain historical context.

2. Environment Generation

Once the knowledge graph is built, the platform constructs the simulation environment.

During this stage, the system performs several tasks:

• entity and relationship extraction


• agent persona generation


• social network construction


• simulation parameter configuration

Agents are assigned identities, backgrounds, and behavioral rules. This ensures that interactions between agents resemble real social dynamics.

3. Parallel Simulation Execution

After the environment is ready, the simulation begins.

Thousands of agents operate simultaneously across the environment, responding to events and interacting with each other. The platform runs simulations across parallel systems, allowing large numbers of agents to operate at the same time.

During this phase the system automatically:

• interprets the prediction request


• simulates social interactions


• updates time-based memory for each agent


• evolves the environment dynamically

The result is a living simulation where narratives, opinions, and behaviors evolve over time.

4. Report Generation

Once the simulation has progressed through multiple cycles, a specialized AI component called ReportAgent analyzes the results.

ReportAgent has access to a rich set of analytical tools and can interact deeply with the simulation environment. It generates a structured prediction report that summarizes:

• key outcomes


• emerging trends


• behavioral insights


• possible risks

This report helps users interpret what happened during the simulation and understand potential real-world implications.

5. Deep Interaction with the Simulation

One of the unique features of MiroFish is that users can interact directly with the simulated world.

Instead of simply reading a prediction report, users can:

• talk with individual agents


• ask questions about their decisions


• explore social dynamics inside the simulation

Users can also communicate with ReportAgent to ask follow-up questions or request deeper analysis.

This interactive layer makes the simulation environment far more flexible than traditional forecasting tools.

Quick Start: Running MiroFish Locally

Developers who want to experiment with the platform can deploy MiroFish locally using either source deployment or Docker deployment.

System Requirements

Before installing the platform, developers need the following tools installed:

To verify installation:

Step 1: Configure Environment Variables

First, copy the example configuration file.

Next, edit the .env file and add the required API keys.

LLM API Configuration

MiroFish supports any LLM API compatible with the OpenAI SDK format.

Example configuration:

The documentation recommends using the Qwen model from Alibaba’s Bailian platform.

Since large simulations can consume significant compute resources, it is recommended to start with simulations of fewer than 40 rounds.

Memory System Configuration

MiroFish uses Zep Cloud to manage long-term memory for agents.

Example configuration:

The free tier of Zep Cloud is usually sufficient for smaller experiments.

Step 2: Install Dependencies

Developers can install all required dependencies with a single command:

Alternatively, the installation can be done step by step.

Install Node dependencies:

Install Python backend dependencies:

This command automatically creates the required Python virtual environment.

Step 3: Launch the Platform

After installation, developers can start both the frontend and backend services with a single command.

Once running, the services are available at:

Frontend interface:

Backend API:

Developers can also start the services separately if needed.

Start only the backend:

Start only the frontend:

Docker Deployment

For teams that prefer containerized environments, MiroFish also supports Docker deployment.

First configure the environment variables as described earlier.

Then start the containers using Docker Compose.

By default, the platform maps the following ports:

• 
  3000 for the frontend interface


• 
  5001 for the backend API

The Docker configuration file also includes commented mirror sources that can be used to speed up container image downloads if needed.

Final Thoughts

While still early in development, swarm intelligence platforms hint at a future where AI systems can simulate complex social environments. Imagine being able to test policies before implementing them, explore market reactions before financial announcements, or examine how information might spread through social networks. Such tools could become powerful decision-support systems for businesses, governments, and researchers. Of course, no simulation can perfectly capture the complexity of real human behavior. Unexpected events and cultural nuances can always influence outcomes.

But platforms like MiroFish show how AI may eventually evolve beyond answering questions and begin modeling entire societies. What began as an experimental open-source project has already sparked significant discussion among developers and researchers. And if multi-agent simulation continues to advance, tools like MiroFish may represent an early step toward a new generation of predictive technologies ones capable of exploring the future inside a digital world before it unfolds in reality.