Ask a what-if question in plain language. MiroFish maps the actors into a knowledge graph, lets a crowd of AI agents live out your scenario on a simulated social surface, and hands back a prediction report you can keep questioning.
Two short recordings from the real product: one full run from question to report, and one session spent interrogating a finished simulation.
Run a product launch or rebrand past a synthetic audience before the real one sees it. The report shows which segment flips first — and in which simulation round — so you know exactly where the message bends.
"If we tease the rebrand a week before the keynote, who leaks, who cheers, and who checks out?"Model how customers, competitors, and commentators respond to a price change. Instead of a single forecast number, you watch objections form, spread, and either fade out or harden into a storyline.
"What do the first 72 simulated hours look like if the pro tier goes up 20%?"Draft a policy or platform-rule change and watch stakeholder groups respond round by round. Useful for spotting which group organizes opposition fastest — and which wording defuses it.
"If the city moves the venue curfew to 11pm, which residents' groups mobilize first?"Track how competing storylines fight for attention after news breaks: which frame wins, who amplifies it, and the round where the tide turns. Built for comms teams deciding whether to respond or wait.
"After the CEO's exit, does 'bold pivot' or 'quiet retreat' dominate by round ten?"Start with a plain-language question, or drop in PDF, Markdown, or text files that describe the situation. The more concrete the actors and incentives in your seed, the sharper everything downstream gets.
MiroFish extracts the people, organizations, relationships, and stakes from your seed and links them into a knowledge graph — the cast list and wiring diagram the simulation will run on.
Each node becomes an autonomous AI agent with its own persona and motives. The agents post, react, argue, and change their minds across rounds on a simulated social surface — so opinion dynamics emerge instead of being scripted.
When the run ends, MiroFish synthesizes what happened into a structured report: the likely outcome, the risk signals worth watching, and the narrative paths the crowd actually took.
The generated world stays alive after the report. Ask it anything — "what if we respond a day earlier?" — and get answers grounded in the simulation you just ran, not a generic model guess.
A condensed sample. Scenario: a consumer brand replaces its longtime human spokesperson with a virtual influencer.
The swap splits the audience by round 4: core fans read it as a betrayal while a younger cohort amplifies the novelty. Net sentiment recovers only if the brand acknowledges the handover directly.
Risk signalsOne decision, one time frame, one audience. "What happens to trial signups in the two weeks after we paywall feature X?" beats "how will people feel about our changes?" — vague questions produce vague crowds.
The best seed files (PDF, MD, or TXT) name specific people and groups and what each one wants. A paragraph like "the moderators fear losing status" does more for the knowledge graph than ten pages of general background.
A prediction report is not a verdict. Mine it for reactions you hadn't considered and questions worth asking next — then test the ones that matter against real-world data before you act.
MiroFish is an open-source multi-agent AI simulation tool for scenario prediction. You describe a situation or upload source material, and it constructs a knowledge graph of the actors involved, runs them as autonomous agents on a simulated social surface, and compiles the outcome into a prediction report you can question afterwards.
It works best on questions where many people react to one event: marketing campaigns, product launches, pricing changes, policy announcements, and public-opinion or market-narrative shifts. If the outcome depends on how a crowd responds, it can be simulated.
No. A plain-language question is enough to start a run. Uploading PDF, Markdown, or text files that name real actors and their incentives gives the knowledge graph more to work with, so the simulation tracks your situation more closely.
An executive summary of the most likely outcome, the risk signals the simulation surfaced, the narrative paths that emerged across rounds, and suggested follow-up questions. After reading it, you can keep chatting with the generated world to probe anything the report left open.
No. MiroFish is exploratory decision support, not a guarantee. Treat a report as a structured rehearsal that surfaces reactions and risks you had not considered — then weigh it alongside your own judgment and real-world data.
Yes. The full source code is public on GitHub at github.com/666ghj/MiroFish. You can read it, run it yourself, file issues, and contribute.