PatronusAI

co-founder @PatronusAI. ex @Meta. prev founded quant fund (acquired in 2020). simulation research ⚡️

Patronus AI is a simulation and evaluation infrastructure company for AI labs and enterprises training agents, and this launch marks its $50M Series B led by Greenfield Partners with participation from Notable Capital, Lightspeed, Datadog, and Samsung, alongside the unveiling of its Digital World Models in limited preview. The product is a class of large-scale simulated digital environments where agents can be trained, evaluated, and stress-tested across long, multi-step workflows before they are exposed to real users. As the team describes it in the launch video, digital world models predict agent states across digital workflows, which lets companies generate the simulation data needed to teach agents the kind of grounding humans accumulate from years of using software. The company was founded in 2023 by Anand Kannappan and Rebecca Qian, who now serve as CEO and CTO respectively, with Anand previously leading responsible AI research and causal inference work at Meta, and Rebecca coming from FAIR, where she trained and released FairBERTa. Patronus first became known for its work on LLM evaluation and benchmarks, and the team has contributed to efforts like Humanity's Last Exam referenced in the launch video. The new round expands that work into reinforcement-learning-driven simulated replicas of websites and internal systems, where agents are iteratively rewarded for successful task completion and penalized for errors. This launch matters now because agent deployments are moving past short prompts into workflows that can run for hours or days, and static benchmarks no longer predict whether those agents will hold up. Patronus is initially focused on verifiable domains like software engineering and finance, with plans to expand into harder-to-verify areas over time , and it positions itself against the internal evaluation teams inside frontier labs rather than human-data vendors, since its approach runs without human reviewers in the loop . For founders and operators building agentic products, it is one of the clearest bets yet on simulation as the training substrate for the next wave of AI systems.

Tags

Comments (15)