A team of researchers at Harvard and MIT has created a virtual world that is big enough for all the inhabitants of the planet. The name of the project is MatrAIx and it relies on 8.3 billion agents which replicate human actions, like talking or buying. Every agent has a specific profile with 1,290 characteristics ranging from shopping preferences to talking patterns.
This idea will allow companies to launch their products without testing them in real environments. The experiments have already taken place 18,000 times and the accuracy of the avatars was at 91.5%.
What is MatrAIx?
MatrAIx is a special kind of simulation platform. A team of 200 scientists developed it in collaboration with researchers from Harvard and MIT and contributors from OpenAI, Anthropic, and Google DeepMind.
The whole endeavor involves use of the dataset called Persona 8B, which contains around 8.3 billion synthetic individuals, representing the current population of human beings closely. Each of these individuals is treated as a persona agent, which means they can react to certain situations and behave like real customers.
When product teams conduct user testing, they usually have to spend months just to recruit testers. But thanks to MatrAIx, they are now able to launch large-scale user research in just hours instead of weeks.
How do persona agents work?
Every agent in the Persona 8B application is composed of 1,290 different attributes, which can be categorized into five main types:
- Background information, which consists of demographic information, such as an individual’s age, residence, and income
- Personality traits, which consists of psychological attributes, such as personality and risk taking
- Job-related capabilities or skills
- Behavioral patterns, which include behavior when shopping online and using apps
- Personal life characteristics, such as daily routines and lifestyle choices
The system generates personas in two different ways. Some personas are constructed based on the dependency graph method, which maintains internal consistency among all the attributes used. Other personas are based on real, anonymized data on humans. There are about 1 million personas in the public version of the system: around 600,000 of them are real personas and 400,000 are purely synthetic.
The behavior simulations of every persona agent come from large language models. Researchers cite GPT-5.5, Claude Opus 4.8, and Claude Haiku 4.5 as simulation models.
What are the four types of user simulation tasks?
MatrAIx evaluates different products in a setting called the MatrAIx Playground that includes four environments applicable in various instances of testing.
| Environment | What it tests |
|---|---|
| Survey | Structured and open-ended feedback for market research and concept testing |
| Chatbot | Task completion, helpfulness, safety, and reliability across conversations |
| Web | Usability, navigation, and task completion on web prototypes |
| App | Functionality, responsiveness, and user preference on app workflows |
As an illustration, if a company wants to assess a newly designed checkout process, thousands of persona agents can go through the Web environment; if one needs help developing a customer support bot, the team can simply put the same agents through the Chatbot environment.
Persona agents vs real human testers: which one fits your project?
The truth is persona agents and Human testers tackle different issues.
Persona agents have their strengths:
- Concept testing during initial product development
- Testing variations cheaply and quickly
- Identifying rare cases that might not be captured with small human samples
Human testers, on the other hand, play an important role in:
- Final assessment before launching the product
- Getting real emotional responses and behavior
- Legal, safety, or major decisions that require accountability
The MatrAIx researchers are honest about the implications of this trade-off. They use the system as a method of discovery and hypothesis formation.
What tools does MatrAIx provide to researchers and product teams?
MatrAIx offers four fundamental components for teams:
- Custom personas – composed profiles that correspond to a predefined target group
- Evaluation infrastructure – platforms and procedures through which the simulations are performed
- Evaluation reports – concise reports on reactions of persona agents
- Telemetry data – unprocessed logs of calls for further research
According to the representative of one audio AI business using the platform, in previous experiments performed prior to employing MatrAIx, the business gave each of its customers the same standard answer.
How can teams start using persona agents today?
MatrAIx has taken the open-source route. The project has birthed its code on GitHub, a community-operated platform, as well as making its one-million-name persona repository available to other researchers.
Becoming acquainted with the initiative follows this procedure:
- Identify the audience for the test
- Choose the perfect mode of interaction: Survey, Chatbot, and other online platforms
- Find or create personas that correspond to the chosen audience
- Carry out the simulation and analyze the assessment report
- Correlate the findings with a limited round of human testing prior to execution
As of now, the platform has capabilities of thousands of application jobs in over 25 different fields of human activity. Including commerce, software production, finance, and healthcare.
Key takeaways
Researchers from Harvard and MIT have developed one of the most significant AI testing tools to date. Enabling them to use 8.3 billion persona agents to represent the population of the world. MatrAIx enables teams to conduct surveys, chatbots conversations, web tests and application tests with synthetic audience. This is before investing in actual testing with real users.
Although, the system does not replace human testing, it enables teams to conduct quick and inexpensive testing. As more teams begin using persona agents for product testing, MatrAIx is expected to become one of the first large-scale and open sources.
