MiroFish: A Swarm Intelligence Engine for Prediction

TL;DR: MiroFish is a swarm intelligence engine that uses multi-agent LLM simulation on top of a knowledge graph to make predictions about complex systems. The same framework has been used for financial forecasting, future-scenario modeling, and event-chain simulation. AGPL-3.0.

Source and Accuracy Notes

• Repository: 666ghj/MiroFish
• License: AGPL-3.0
• Stars: 73,191 (as of September 2026)
• Language: Python
• Topics: agent-memory, financial-forecasting, future-prediction, knowledge-graph, llms
• Homepage: mirofish.ai

What It Is

MiroFish is a Python framework for running multi-agent simulations over a knowledge graph. Instead of asking a single LLM “what do you think will happen,” the engine seeds an environment with autonomous agents that interact, exchange information, and collectively produce forecasts or scenarios.

The architecture has two halves:

• A knowledge graph layer that stores facts, entities, and relationships extracted from input data — news, filings, public posts, historical events. Agents read and write to this graph as they simulate.
• A swarm simulation layer where each agent runs an LLM-driven policy, observes the world, communicates with neighbors, and updates its beliefs.

The result is a system that can produce not just a single point forecast but a distribution of plausible futures, with traces of how each emerged from the underlying agent interactions. This makes it especially useful for forecasting problems where causality is opaque and second-order effects matter.

Why It Matters

Single-prompt “ask an LLM to predict X” is brittle. The model is effectively doing one inference pass against its training distribution and producing a hallucinated average. For domains like financial markets, supply chains, or geopolitics — where the answer depends on how many actors respond to each other — that approach is structurally weak.

MiroFish’s swarm-on-graph approach addresses several of those weaknesses:

• Emergent dynamics. Multiple agents with different roles can model how a central bank’s move ripples through markets, or how a product launch ripples through communities.
• Traceable reasoning. Every prediction comes with a log of which agents said what, and how the graph evolved. You can audit the chain.
• Domain-agnostic. The same engine has been applied to financial forecasting, future-prediction problems, and complex event simulation. You’re not locked into one vertical.
• Reproducible runs. Given fixed seeds and graph inputs, the same simulation can be replayed.

The knowledge graph also doubles as a long-term memory layer. Past simulations can be queried, replayed, and compared, which is useful for tracking how predictions evolve as new evidence arrives.

Practical Trade-Offs

A few honest constraints:

• AGPL-3.0. This is a strong copyleft license. Any service that uses MiroFish and offers it over a network must publish its source. For internal research use it’s fine; for commercial SaaS you need a separate license.
• Compute cost. Running dozens of LLM agents in parallel is expensive. A single simulation can easily cost tens of dollars in API calls. This is not a real-time tool.
• Garbage in, garbage out. The quality of the knowledge graph input dominates everything else. If your data ingestion pipeline is sloppy, the simulation will be confidently wrong.
• Calibration is your job. MiroFish produces plausible narratives. Translating those into calibrated probabilities or actionable forecasts requires a separate evaluation layer.

These caveats are not unique to MiroFish — they apply to any LLM-driven simulation framework — but they shape what the tool is good for.

Who Should Care

MiroFish is worth a serious look for:

• Financial researchers modeling market scenarios beyond naive trend-following.
• Geopolitical and policy analysts who want to explore second-order effects of events.
• Strategy teams running red-team or war-game exercises against a dynamic environment.
• Academic researchers studying multi-agent LLM systems, complex systems, or forecasting tournaments.

It’s less suited for anyone who needs a single deterministic number — a regression model will beat it there. Where it shines is producing structured, traceable scenarios in domains where naive forecasting has already failed.

Summary

MiroFish is one of the more ambitious open-source multi-agent simulation frameworks available. It’s not a forecasting oracle, and it shouldn’t be sold as one. As a research tool for exploring complex systems with structured, traceable multi-agent simulation, it’s a substantive piece of infrastructure — provided you’re comfortable with AGPL and the compute bill.