MatrAIx

MatrAIx is a research project and service that builds a corpus of 8.3 billion simulated persona agents, each with its own preferences, routines and digital context. It began as a way to test product changes (A/B tests) with persona agents before real users see them, and has since been extended to evaluating personal agents.

The offering consists of persona-based simulated users, evaluation infrastructure for running an agent across tasks, domains and environments, results and analysis broken down by user, domain and failure mode, persona-conditioned SFT trajectories, and RL environments with simulated users. Related benchmarks listed on the site include PersonaBench (AgentPersonaBench across CUA, chat, code, science and robotics) and PersonalAgentEval, which covers seven commercial personal agents, 152 tasks in seven domains and simulated users with calendars, inboxes, orders and health records. In the published PersonalAgentEval results, the top-scoring agent (Instinct) resolved 60.5% of tasks.

Access appears to be through the vendor, aimed at teams building personal agents; no public pricing, code or model downloads are listed on the pages provided. A YouTube reviewer noted it is unclear how closely the simulated data matches real people's behavior.