How Can I Test My Software With Thousands of Users? Use AI Agent Simulation with MiroFish
MiroFish is an open-source multi-agent AI simulation engine that creates a virtual world containing thousands of AI agents with different personalities, memories and behavioral patterns. It can simulate how these agents may react to policies, products, events, content or other scenarios and generate prediction reports. This makes it potentially useful for testing product ideas, user reactions, safety scenarios, marketing strategies and other “what-if” situations before applying them in the real world.
How Can I Test My Software With Thousands of Users?
Testing software with thousands of users can be difficult when you do not yet have thousands of real users available. Traditional user testing requires recruiting people, creating test scenarios, collecting feedback, and analyzing their behavior.
AI-powered multi-agent simulation provides another approach.
MiroFish is an open-source multi-agent AI prediction engine that can create a simulated digital environment containing thousands of AI agents with different personalities, memories, and behavioral patterns.
Instead of asking one AI model what users might do, MiroFish can simulate interactions between many AI agents and analyze the resulting behavior.
Important: MiroFish simulates thousands of AI agents. These are not thousands of real human users or real customers.
What Is MiroFish?
MiroFish is a swarm-intelligence and multi-agent simulation engine designed to simulate possible future scenarios.
You provide seed information, such as documents or other scenario data, together with a natural-language prediction requirement. MiroFish then builds a simulated environment where AI agents can interact and evolve.
The official workflow consists of:
- Graph Building
- Environment Setup
- Simulation
- Report Generation
- Deep Interaction
The system uses knowledge graphs, GraphRAG, agent personas, memory, and multi-agent simulation to create the simulated environment. :chatgpt-content-reference{index="1"}
Can MiroFish Simulate Thousands of Users?
Yes, the project's core concept is to create a parallel digital world containing thousands of intelligent agents.
However, these should be understood as virtual users or simulated AI agents, not actual people using your software.
For software testing and product research, this distinction is important.
MiroFish can help answer questions such as:
- How might different types of users react to a new feature?
- What problems could users encounter in a workflow?
- What objections might users have?
- How could different user personas behave?
- What could happen if we change a product workflow?
- Which potential outcomes should we investigate with real users?
The MiroFish repository specifically describes its agents as having independent personalities, long-term memory, and behavioral logic. :chatgpt-content-reference{index="2"}
How Does AI User Simulation Work?
A simplified MiroFish workflow looks like this:
Your Software or Scenario → User Data/Scenario → Knowledge Graph → AI Personas → Multi-Agent Simulation → Behavior Analysis → Prediction Report
1. Provide the Scenario
Start with information describing the product, feature, workflow, policy, or situation you want to test.
For example:
We are launching a new AI-powered safety-management workflow for construction companies.
You could provide supporting documentation explaining:
- The product
- The feature
- Existing workflows
- User roles
- Business rules
- Expected behavior
- The problem being solved
MiroFish accepts seed materials and a natural-language prediction requirement. :chatgpt-content-reference{index="3"}
2. Build a Knowledge Graph
MiroFish extracts entities and relationships from the provided information and uses GraphRAG to construct a representation of the scenario.
This provides the simulated agents with contextual information about the environment they are operating in.
3. Generate Different User Personas
The system can create different agent personas with different characteristics.
For software testing, you could model personas such as:
- New user
- Experienced user
- Administrator
- Manager
- Employee
- Technical user
- Non-technical user
- Highly engaged user
- Resistant user
- Time-constrained user
The goal is to avoid testing a product against one generic AI personality.
Instead, different agents can represent different behavioral patterns.
4. Run the Simulation
The agents then interact within the simulated environment.
MiroFish's workflow includes multi-agent simulation and dynamic temporal memory updates. The project also describes deep interaction with individual agents after simulation. :chatgpt-content-reference{index="4"}
This makes it possible to explore how different simulated users could react to the same scenario.
5. Analyze the Results
After the simulation, MiroFish generates a prediction report based on the simulated environment.
The report can help identify:
- Common behavioral patterns
- Potential problems
- Different user reactions
- Possible risks
- Emerging trends
- Unexpected outcomes
- Areas requiring further testing
How Can I Use MiroFish to Test My Software?
MiroFish can potentially be used as an additional layer of product research and scenario testing.
Software Feature Testing
Suppose you are launching a new feature.
Instead of immediately releasing it to everyone, you could define a simulation such as:
Simulate how different types of SaaS users may react to this feature and identify potential usability problems.
The simulated population could contain different user personas.
The output could help the product team identify questions that should be validated with real users.
UX and Workflow Testing
You can simulate workflows such as:
- Registration
- Onboarding
- Creating records
- Completing forms
- Managing tasks
- Generating reports
- Approving requests
- Completing safety workflows
- Managing employees
- Using dashboards
For example:
Simulate how a new employee with limited technical knowledge would attempt to complete this workflow.
You could then compare the simulated behavior with the intended user journey.
Product Adoption Simulation
MiroFish can also be considered for exploring product adoption scenarios.
For example:
Simulate how different customer personas might respond to the introduction of an AI assistant.
You could investigate:
- Who might adopt it quickly?
- Who might resist it?
- What concerns could appear?
- What features might encourage adoption?
- What questions could users ask?
These results should be treated as hypotheses rather than statistically proven customer behavior.
Example: Testing a SaaS Product With 1,000 AI Users
Imagine you have a SaaS application with a new dashboard.
You want to understand how different users may react.
You could define personas such as:
- 200 administrators
- 300 managers
- 400 employees
- 100 power users
Each persona can have different characteristics and objectives.
The simulation could then explore scenarios such as:
"Introduce the new dashboard and simulate how different user personas interact with it. Identify potential confusion, feature preferences, workflow problems, and likely adoption barriers."
The resulting simulation could reveal patterns worth investigating before conducting a real-world user test.
Testing Safety Management Software
MiroFish could be particularly interesting for safety-management software because many workflows involve different users and decision-making scenarios.
For example, consider a construction safety platform with:
- Risk assessments
- TRA
- LMRA
- Training
- Certificates
- Findings
- Corrective actions
- Work permits
You could create a simulation around a scenario such as:
"A construction company introduces a mandatory LMRA workflow before high-risk work begins. Simulate how employees, supervisors, and safety managers could respond."
Potential areas to explore include:
- Employee compliance
- Supervisor reactions
- Workflow friction
- Communication problems
- Possible resistance
- Different interpretations of safety requirements
- Potential improvement opportunities
This does not replace actual safety testing or real-world validation. It provides an additional way to explore possible scenarios.
Testing Marketing and Product Ideas
The same concept can be applied outside traditional software testing.
For example, you could simulate reactions to:
- A product announcement
- A new pricing model
- A landing page
- A marketing campaign
- A new product feature
- A policy change
- A new AI capability
A simulation could help identify possible:
- Customer objections
- Questions
- Positive reactions
- Adoption barriers
- Messaging opportunities
- Unexpected responses
MiroFish's own examples include public-opinion simulation and other future-prediction scenarios. :chatgpt-content-reference{index="5"}
What MiroFish Does Not Do
MiroFish should not be described as a tool that gives your software thousands of real users.
It does not automatically provide:
- 1,000 real customers
- 1,000 real testers
- Real production traffic
- Guaranteed product-market validation
- Guaranteed predictions of customer behavior
Instead, it provides a way to create simulated AI agents that can represent different possible users and behaviors.
The distinction is:
Real User Testing
Real people → Real software → Real behavior → Real feedback
AI Agent Simulation
AI personas → Simulated environment → Simulated behavior → Hypotheses and predictions
The strongest approach can be to use both.
AI Simulation + Real User Testing
A practical product-development workflow could be:
Step 1: Use MiroFish to explore thousands of simulated user scenarios.
Step 2: Identify the most interesting risks, questions, and behavioral patterns.
Step 3: Select the most important scenarios.
Step 4: Test those scenarios with real users.
Step 5: Compare real-world results with the simulation.
Step 6: Improve the simulation using real-world feedback.
This creates a feedback loop between AI simulation and real user research.
Why Use AI Agents Instead of One AI Prompt?
A single AI prompt might answer:
"How would users react to this feature?"
But the answer is still generated from one model interaction.
A multi-agent simulation takes a different approach.
Different agents can have different:
- Personas
- Memories
- Opinions
- Behavioral patterns
- Relationships
- Interactions
Those agents can then interact with one another inside a simulated environment.
MiroFish is specifically designed around this multi-agent approach. :chatgpt-content-reference{index="6"}
MiroFish Workflow
The project's documented workflow is:
1. Graph Building
Seed information is extracted and used to build a knowledge graph and memory structure.
2. Environment Setup
Entities, relationships, personas, and agent configurations are prepared.
3. Simulation
Agents interact within the simulated environment and their temporal memories can change.
4. Report Generation
A ReportAgent analyzes the simulation and generates a prediction report.
5. Deep Interaction
Users can interact with simulated agents and explore the simulated environment after the simulation. :chatgpt-content-reference{index="7"}
What Do You Need to Run MiroFish?
The official repository currently lists:
- Node.js 18+
- Python 3.11–3.12
- uv package manager
- An LLM API compatible with the OpenAI SDK format
- Zep Cloud configuration
The project also supports Docker deployment. :chatgpt-content-reference{index="8"}
The official documentation recommends starting simulations with fewer than 40 rounds because of potentially high LLM consumption. :chatgpt-content-reference{index="9"}
Important Limitations
AI simulation is not a replacement for real user testing.
Simulation results depend on:
- Quality of the source information
- Persona generation
- LLM capabilities
- Simulation configuration
- Number of agents
- Number of simulation rounds
- Quality of the underlying assumptions
A simulated agent is not equivalent to a real person.
Therefore, simulation results should be treated as research signals and hypotheses, not guaranteed predictions.
For important product decisions, simulated results should be validated with real users and real-world data.
Is MiroFish Worth Exploring for Software Testing?
If your goal is to explore how large groups of different user types might react to a product, feature, workflow, policy, or scenario, MiroFish is an interesting technology to investigate.
Its biggest potential is not simply "testing with thousands of users."
It is the ability to create a virtual population of AI agents and explore many possible behavioral outcomes before spending the time and money required for large-scale real-world testing.
A useful way to think about it is:
MiroFish gives product teams a virtual laboratory for exploring what thousands of simulated users might do.
Final Takeaway
If you are asking:
"How can I test my software with thousands of users when I don't have thousands of users?"
AI multi-agent simulation is one approach worth exploring.
MiroFish provides an open-source framework for creating simulated environments containing thousands of AI agents, generating different personas, running multi-agent simulations, and producing prediction reports. :chatgpt-content-reference{index="10"}
It should not replace real customers or real usability testing, but it can potentially help product teams discover questions, risks, behavioral patterns, and scenarios that deserve real-world validation.
In short:
Thousands of simulated AI users → Explore scenarios → Find patterns → Validate with real users → Improve the product