MiroFish — Predicting the Future by Simulating Thousands of AI Agents

Digital Petri Dishes

What happens when you release thousands of AI agents into a digital world and let them interact? MiroFish starts from this question.

Academic Roots — Stanford Generative Agents

MiroFish didn't emerge from nowhere. The 2023 "Generative Agents" paper by Joon Sung Park's team at Stanford was the direct inspiration.

In that study, researchers placed 25 AI agents in a small virtual town, giving each a name, occupation, personality, and relationships, then letting them interact freely. The results were remarkable:

• Agents spontaneously organized a party and distributed invitations
• Agents with similar interests naturally formed groups
• The way rumors spread about other agents matched real sociological theories

MiroFish scales this experiment from 25 to thousands or tens of thousands of agents and turns the resulting patterns into a prediction tool.

How It Works — Detailed Architecture

Traditional ML learns patterns from historical data. MiroFish takes a different path. It gives LLM-powered agents distinct personalities, memories, and behavioral patterns, then lets them interact in simulated environments.

Agent Structure

Each agent consists of these modules:

• Persona Module: Personality traits, values, decision-making tendencies (based on Big Five personality model)
• Memory Stream: Stores experiences chronologically with importance-weighted retrieval priorities
• Reflection Engine: Automatically extracts high-level insights from accumulated experiences (runs reflection every 50 memories)
• Planning Module: Auto-generates and adjusts long-term goals and short-term action plans
• Social Graph: Dynamically tracks relationship strength and nature with other agents

Simulation Process

1. World Building: Define the environment and rules to simulate (market, city, organization, etc.)
2. Agent Seeding: Assign initial characteristics and goals to each agent
3. Interaction Loop: Agents freely interact (conversations, transactions, cooperation, conflict)
4. Emergent Pattern Detection: Statistically analyze patterns that emerge from collective behavior
5. Prediction Extraction: Forecast future trends based on detected patterns

Cost Optimization

Running thousands of agents on LLMs could be astronomically expensive. MiroFish solves this:

• Tiered Model Routing: Simple actions use small models (Phi-3, Gemma 2), complex decisions use large models (Claude, GPT-4o)
• Batch Processing: Concurrent agent actions are batched together
• Cached Personas: Repetitive personality patterns are cached to minimize LLM calls
• Result: 1,000 agents x 100 timestep simulation costs approximately $15-30

Real-World Applications

1. Financial Market Simulation

Give agents investor profiles (conservative, aggressive, momentum-following, etc.) and let them trade in a virtual market. Simulate market reactions when new policies are injected (interest rate hikes, regulatory changes).

One hedge fund reportedly predicted Japanese market volatility 5 days ahead during beta testing of August 2025 (unofficial).

2. Product Launch Simulation

Assign consumer agents various demographic characteristics and purchasing tendencies, then simulate reactions to a new product launch. Pre-test price elasticity, channel effectiveness, and competitive positioning.

3. Policy Simulation

Simulate how citizen agents change behavior when new transportation policies are introduced to a city. Pre-analyze effects on public transit usage, traffic congestion, and real estate prices.

4. Social Media Viral Prediction

Give agents social media user characteristics and simulate how specific content spreads. Analyze influencer effects and algorithmic change impacts.

GitHub Growth Data

• GitHub Trending #1 (as of March 7)
• 28,600 stars (current)
• +2,782 stars in 24 hours
• Key contributors: 12 (3 core contributors are former Google DeepMind researchers)
• License: Apache 2.0

Comparison with Existing Simulation Tools

Tool	Approach	Agent Scale	LLM-Based	Cost
MiroFish	LLM agent simulation	Thousands-tens of thousands	Yes	$15-30/run
NetLogo	Rule-based ABM	Tens of thousands-millions	No	Free
Mesa (Python)	Rule-based ABM	Thousands-tens of thousands	No	Free
HASH	Large-scale simulation	Millions	No	Paid

MiroFish's key differentiator: Rule-based ABM (Agent-Based Models) require pre-programming agent behavior — "if price rises 10%, demand drops 5%." These rules are defined by humans. MiroFish uses LLMs that understand situations and autonomously decide actions, enabling unexpected emergent behaviors.

Limitations and Caveats

• Hallucination risk: LLM-based agents can exhibit unrealistic behaviors
• Validation difficulty: Verifying how well simulation results reflect reality is inherently challenging
• Cost: Still more expensive than rule-based ABM (but dropping rapidly)
• Reproducibility: Due to LLM's stochastic nature, the same simulation produces different results each time. Ensemble approaches are necessary

Related Projects and Background

A Brief History of Agent-Based Modeling (ABM)

ABM originated in the 1970s with Thomas Schelling's segregation model. A simple rule — "move if fewer than 30% of neighbors share your race" — produced extreme segregation as an emergent phenomenon. It evolved through Santa Fe Institute's complexity research, Epstein & Axtell's "Sugarscape" (1996), and tools like NetLogo (1999). MiroFish combines this 50-year ABM tradition with LLMs.

Simulacra and Other LLM Simulation Projects

Several similar projects emerged after Stanford's Generative Agents paper:

• CAMEL (2023): Task execution through role-playing between LLM agents
• MetaGPT (2023): Multi-agent simulation of a software company
• AgentVerse (2024): General-purpose multi-agent simulation framework
• MiroFish (2026): Specialized for large-scale prediction from simulations

MiroFish's differentiator: simulation isn't the goal — extracting patterns from simulations to predict reality is.

Connection to Computational Social Science

In the broader picture, MiroFish is an extension of computational social science. Duncan Watts (Microsoft Research) argued in a 2007 Nature paper that "social phenomena can't be reproduced in laboratories, making simulation essential." If LLMs can approximate human decision-making, social simulation realism improves dramatically. A 2025 Nature study found that LLM agent collective behavior patterns showed 72% correlation with actual human experimental data.

Why It Matters

Agent-based simulation is one of AI's next frontiers. Instead of single-model reasoning, it extracts insights from collective behavior of many agents. High potential for finance, policy, marketing, and social phenomenon prediction.

What's even more exciting: this kind of simulation is being democratized. Social simulations that once required RAND Corporation or major consulting firms can now be run by a single developer on a laptop.