Multi-agent AI prediction engine
One event. Thousands of perspectives.
A world of possible outcomes.
Turn real-world information into a society of AI agents. Explore how people, markets, and ideas might respond—before a decision leaves the drawing board.
Open-source foundation. Human judgment at the center.
Individual reactions travel through relationships. A shared story can become a collective response.
Conceptual illustration · not a simulation result01 / The idea
A digital sandboxMiroFish is an open-source, AI-powered prediction engine that uses multi-agent swarm intelligence to explore how complex situations could unfold.
Give it the ingredients of a situation: a news article, a policy proposal, financial signals, research, or even a fictional world. MiroFish uses that seed information to construct a virtual society. Its agents represent different viewpoints, with their own personalities, memory, and behavioral logic. As they interact, the simulation can reveal tensions, alliances, shifting narratives, and unexpected consequences.
The important unit is the interaction. A customer responds to a price change; a competitor responds to the customer; a news story reframes the debate. Instead of stopping at an initial answer, a multi-agent simulation explores how one reaction can change the next. This makes it useful for asking better questions about markets, public policy, business decisions, and social dynamics.
The project describes environments with thousands of autonomous agents and an interactive, high-fidelity digital world. Actual scale and fidelity depend on the chosen setup, models, source material, and available compute. A more detailed simulation can represent more relationships, but detail alone does not establish accuracy.
The result is a prediction report and a world you can investigate. Treat them as a way to compare possibilities, identify assumptions, and prepare research—not as a promise that events will happen. A useful simulation helps you find the next decision to test in reality.
02 / Inside the engine
Four stages connect your question to a world of responses. The quality of the question matters at every step.
Start with news, policy drafts, financial reports, research notes, or a scenario written in natural language. Define the decision, the people affected, and the time horizon.
Make it concreteName what is known, what is uncertain, and which assumptions the simulation should hold constant.
INPUT → Context & a questionThe engine organizes entities and relationships, creates agent personas, and sets up the simulation environment. Agents bring different incentives, memories, and ways of responding.
Let interactions matterPeople can respond to the event and to one another. Discussion, disagreement, and feedback shape what follows.
MODEL → Agents & relationshipsExplore the project’s “God’s-eye view” idea: introduce a variable or revise a scenario, then investigate how reactions change. Compare a baseline with a different decision.
Ask a sharper what-ifWhat if a policy starts later? What if a competitor cuts prices? Change one assumption at a time when comparing runs.
EXPLORE → Alternative pathsA prediction report brings together the simulation’s patterns, stakeholder reactions, and possible outcomes. Investigate the reasoning and follow up with agents or the report assistant.
Turn findings into questionsLook for fragile assumptions, minority views, and evidence that would change your decision.
OUTPUT → A report to investigateThe upstream workflow includes graph building, environment setup, simulation, report generation, and follow-up interaction. Read the project workflow ↗
03 / Collective intelligence
A virtual society is more than a collection of separate AI answers. Its value comes from watching perspectives affect each other.
What changes when everyone else starts reacting, too?
The question behind swarm intelligenceDifferent roles and incentives create different reactions to the same event. A customer, regulator, employee, and investor do not start with the same priorities.
Past interactions can inform later behavior. This makes it possible to explore how trust, repeated messages, or earlier disagreements influence the next round.
Individual exchanges can lead to collective patterns: agreement, polarization, new alliances, or a narrative that spreads beyond its original audience.
Follow the relationships behind a conclusion. The project supports interaction with simulated agents and its ReportAgent after a run.
Explain the question in ordinary language. Running a self-hosted installation still requires technical setup, but describing the scenario does not require a modeling language.
04 / Where to use it
These are starting points for exploration. Each example is a question to investigate, not a forecast produced by this website.
Explore how investor sentiment, company announcements, supply shocks, or changing expectations could interact. Map competing narratives and the assumptions behind a bullish, bearish, or uncertain response.
Useful inputs: Earnings commentary, public news, sector context, and an explicit time horizon. Questions to bring back: Which signal drives the reaction? What would make the narrative reverse?
“How might retail investors, analysts, and suppliers respond if a company reduces its revenue outlook?”
Investigate possible responses to a proposed law, a transport policy, a public announcement, or a geopolitical event. Identify where stakeholder goals conflict and how communication could change the discussion.
Useful inputs: The proposal, affected groups, local context, and alternative implementation plans. Questions to bring back: Who bears the cost? Which concerns need real consultation?
“How could residents, commuters, and local shops react to a city-center congestion charge?”
Rehearse a pricing change, product launch, market entry, or marketing campaign. Explore the second-order effects: a new offer changes expectations, competitors react, and existing customers reconsider their choices.
Useful inputs: Product facts, customer segments, competitive context, and the decision being tested. Questions to bring back: Which objection is easiest to miss? Which alternative deserves a customer interview?
“What could happen if a subscription product raises prices while introducing a lower-cost entry plan?”
Give characters different motivations and explore where their choices could lead. Use a story premise, a fictional society, or an alternate-history question to investigate plausible tensions and branching storylines.
Useful inputs: Character backgrounds, world rules, earlier events, and the point of divergence. Questions to bring back: Does the outcome respect each character’s motivation? What surprising conflict would improve the story?
“How would a coastal town change if its residents discovered a reliable way to communicate with the future?”
05 / Think in alternatives
A useful what-if exercise makes its assumptions visible. Start with a baseline, change one meaningful condition, and compare what you would look for.
SCENARIO DESIGN EXAMPLE
Choose an assumption to see how to frame the research. These are authored planning prompts, not AI-generated predictions.
The proposed area, price, affected groups, and observation period.
The charge begins on a single announced date, without a transition period.
How commuters describe the immediate cost, how shops respond to perceived footfall changes, and where exceptions become contentious.
Compare the discussion with travel data, consultations, and evidence from comparable cities.
06 / After the simulation
The most useful result is a clearer picture of what to investigate. Read the report alongside the scenario, source material, and the interactions that led to it.
Explore the upstream demoA READER’S GUIDE TO THE OUTPUT
What was asked, which information shaped the world, and what the simulation took for granted.
Where groups agree, where they diverge, and which less-visible perspectives challenge the main story.
Recurring narratives, feedback effects, possible turning points, and outcomes worth comparing.
Uncertain assumptions, unanswered questions, and research or experiments that could support a real decision.
Use this as a reading checklist. The contents and format of a generated report depend on the scenario and implementation.
07 / A different way to ask
These approaches complement each other. Choose the one that matches the question you need to answer.
| What you explore | A single AI response | Multi-agent simulation |
|---|---|---|
| Perspective | One synthesized response to a prompt. | Different roles, incentives, and viewpoints. |
| Change over time | A description of what might happen. | A sequence of reactions and interactions. |
| What-if questions | Ask for a revised answer. | Change the scenario and compare behavior. |
| Investigation | Follow up on the explanation. | Explore the report and individual agents. |
| Cost & complexity | Usually lighter and faster. | More model calls, setup, and analysis. |
08 / Use it thoughtfully
MiroFish is useful when you want to challenge assumptions, explore reactions, or compare a few plausible paths. It is less suitable when you need a verified probability, a precise stock-price target, or a substitute for real people’s views.
A virtual population is a model. Agent personas can miss cultural context, underrepresented voices, and behavior that was absent from the seed data.
Consensus is not confirmation. Many agents can share the same underlying model assumptions. Agreement inside a simulation does not make a claim true.
More runs are an exploration tool. Repeated simulations help investigate sensitivity; hundreds or thousands of runs do not automatically yield calibrated forecasts.
Use evidence beyond the sandbox. Validate ideas with current data, interviews, experiments, and subject-matter expertise. Do not use a simulated market scenario as an investment recommendation.
09 / Open by design
MiroFish’s source is available on GitHub under the AGPL-3.0 license. You can inspect the implementation, explore the project’s examples, and follow its development.
See the upstream project’s prepared environment before deciding whether to run your own installation.
View the public demo ↗The documented source setup uses Node.js, Python, model API access, and Zep Cloud. Open-source code does not mean model calls or infrastructure are free.
Check current requirements ↗MiroFish.shop is an independent introduction to the open-source project. This website explains the engine and links to upstream resources; it does not run simulations or represent the original project team.
10 / Good questions
A few useful distinctions for your first exploration.
The upstream project describes a virtual society containing thousands of AI agents. Each agent represents a persona with its own context and behavior. The actual number in a run depends on configuration and resources; it is not a claim that this informational website is running thousands of agents.
Start with relevant source material and a specific natural-language question. Identify the decision, affected stakeholders, known facts, assumptions, and time horizon. Separate facts from speculation so you can interpret the output in context. Use only information you have permission to share with your chosen deployment.
Scenario exploration is central to the project’s concept. Compare a baseline with a changed variable, such as the timing of a policy or a competitor’s response. The controls available during or between runs depend on the implementation. Preserve the same starting assumptions where possible so comparisons remain meaningful.
No accuracy rate or guaranteed outcome is claimed here. Simulation results depend on source quality, agent design, model behavior, and the assumptions you choose. A plausible narrative is not a validated forecast. Use results to guide investigation and test them against observed evidence.
You can describe scenarios in natural language. Running the open-source version yourself requires installing and configuring its services and API connections. The project’s prepared demo is an easier starting point for exploring the experience.
The upstream code is available under AGPL-3.0. Hosting, model API calls, and connected services may have separate charges. Review the license and current setup requirements before choosing a deployment. This website does not sell a subscription or process payments.
A chatbot can quickly synthesize an answer or discuss alternatives. MiroFish’s multi-agent approach models different participants reacting to a situation and to one another over time. That interaction is useful for exploring social dynamics, at the cost of more computation and more assumptions to examine.
This site is a detailed guide, not a hosted prediction service. To explore a prepared simulation, open the upstream demo. To configure your own environment, follow the project setup guide.
The future starts with a better question.
MiroFish