Graph Building
The workflow extracts seed information, builds ontology and relationships, and prepares memory that later agents can use.
AI simulator / MiroFish internals
An AI simulator like MiroFish works by turning source material into agent behavior.
MiroFish is an AI simulator for scenario rehearsal: upload seed material, build a graph, configure agent personas, run interaction rounds, generate a report, and continue deep interaction with agents or the report.
AI simulator
Graph Building
Agent setup
ReportAgent
How the AI simulator works
The MiroFish AI simulator follows a five-stage source-to-report loop.
The official MiroFish README describes the project as a swarm intelligence engine for prediction. It says users upload seed materials and describe prediction requirements in natural language, then MiroFish returns a detailed prediction report and an interactive digital world.
The same open-source README names the workflow stages: Graph Building, Environment Setup, Simulation, Report Generation, and Deep Interaction. Those stages are visible in the public source structure through frontend components such as Step1GraphBuild.vue, Step2EnvSetup.vue, Step3Simulation.vue, Step4Report.vue, and Step5Interaction.vue.
In plain terms, the system does not begin by guessing. It first organizes source material into entities and relationships, prepares agent personas, lets those agents act in a simulated environment, then uses the post-simulation state to generate a report the user can question.
Five stages
Workflow from open-source MiroFish
Environment Setup
The setup step turns graph entities into personas, simulation configuration, time windows, and behavior settings.
Simulation
MiroFish runs parallel interaction rounds so posts, comments, searches, votes, or reactions can evolve over time.
Report and interaction
The report step summarizes the run, then lets the user ask ReportAgent or simulated agents follow-up questions.
Main functions of the MiroFish workflow
The first function is source grounding. A useful run begins with files, notes, reports, stories, or event briefs. The source packet gives the simulation something to respect.
The second function is actor modeling. MiroFish needs people, groups, roles, incentives, constraints, and relationship context. Without actors, the report becomes a generic essay.
The third function is controlled interaction. MiroFish does not only summarize the brief. It runs simulated perspectives through rounds so disagreements, shifts, and fragile assumptions become visible.
The fourth function is report review. The output should show what happened in the simulated world, why a path appeared, what evidence is weak, and what the user should test next.
AI simulator pseudo-code for the MiroFish mental model
const mirofishAiSimulator = {
input: ['source files', 'scenario requirement', 'actors', 'constraints'],
stages: [
'Step1GraphBuild.vue: build ontology, entities, relationships, memory',
'Step2EnvSetup.vue: create personas and simulation configuration',
'Step3Simulation.vue: run dual-environment interaction rounds',
'Step4Report.vue: generate a reviewable prediction report',
'Step5Interaction.vue: ask ReportAgent or simulated agents follow-up questions'
],
output: ['reaction branches', 'assumptions', 'weak signals', 'next tests']
}This pseudo-code is not copied from the repository. It is a compact map of the AI simulator workflow described by the public README and component names. Use it as a reading guide before opening the source.
Source-code reading path
Start with the UI stages because they show the user journey in order. GraphPanel.vue helps you inspect entities and relationships. Step1GraphBuild.vue explains the graph-building entry. Step2EnvSetup.vue shows how personas and simulation settings become visible. Step3Simulation.vue is the active run stage. Step4Report.vue connects the run to a report. Step5Interaction.vue turns the report into follow-up conversation.
Then read the backend and runtime patches only after the user flow is clear. In this site repository, runtime-server and test fixtures also mention Step3Simulation.vue, SimulationManager, stop_simulation_for_report, and ReportAgent. Those names support the same source-to-report mental model without requiring this visitor page to become a full developer manual.
For a product evaluator, those landmarks answer a practical question: where does the workflow move from material, to world model, to behavior, to report? If those transitions are clear, the visitor can judge whether MiroFish fits their scenario before reading deployment details.
Operating steps for a first MiroFish run
Prepare one source packet
MiroFish needs a focused packet: a report, policy draft, product page, event brief, story chapter, or market note.
Write a testable requirement
The workflow works better when the requirement names the outcome, time horizon, actors, constraints, and the decision being considered.
Review the graph and personas
Before simulation, inspect whether key entities, relationships, and generated agents match the source material.
Run a small pilot first
The upstream README warns that model use can be expensive and recommends trying fewer than 40 rounds first.
Read the report with skepticism
A report is useful when it names assumptions, simulated events, and the follow-up evidence that should be checked.
Rerun with one changed variable
Change one assumption, message, price, event, or constraint so the difference is easy to interpret.
AI simulator evidence table
| Stage | What the user should inspect | Why it matters | Common mistake |
|---|---|---|---|
| Graph Building | entities, relationships, source coverage | bad graphs create bad agent context | accepting irrelevant entities |
| Environment Setup | personas, time windows, active agents, behaviors | MiroFish needs plausible roles | running with default settings blindly |
| Simulation | event feed, round progress, action mix, repeated behavior | the report depends on simulation quality | treating an unfinished run as final |
| Report Generation | assumptions, traceable claims, weak evidence, branch differences | reviewability beats confident prose | quoting generated claims as facts |
| Deep Interaction | follow-up questions, agent reasoning, one-variable reruns | the AI simulator becomes useful through iteration | asking broad new questions too soon |
Start from MiroFish
The homepage shows the AI simulator media, product path, guides, and pricing route.
Use this guide to understand how MiroFish works. Use the homepage when you are ready to inspect the full product experience.
FAQ
AI simulator FAQ
What is the MiroFish AI simulator workflow?
The public MiroFish workflow has five stages: Graph Building, Environment Setup, Simulation, Report Generation, and Deep Interaction.
Which source components show the AI simulator stages?
The frontend component list includes GraphPanel.vue, Step1GraphBuild.vue, Step2EnvSetup.vue, Step3Simulation.vue, Step4Report.vue, and Step5Interaction.vue.
What should a first AI simulator run use?
Use one focused source packet, a clear scenario requirement, named actors, a short time horizon, and a review plan.
Source: Official 666ghj/MiroFish README and frontend component directory.
Method: Summarized public workflow stages, then mapped them to practical user checks.
Limits: Component names and implementation details can change; use current upstream source for final code review.
Updated: July 17, 2026