Digital world models that simulate software environments for training and stress-testing AI agents

Patronus AI was founded in 2023 by Anand Kannappan and Rebecca Qian, both previously machine learning researchers at Meta, and is based in San Francisco. The company started in LLM evaluation, shipping automated scoring for hallucination and unsupported claims and releasing open research artifacts including the Lynx hallucination detection model, the GLIDER explainable judge, and FinanceBench, a benchmark for financial question answering. It raised a $17M Series A led by Notable Capital in 2024, with Lightspeed, Datadog, and others participating, and later announced a $50M Series B led by Greenfield Partners, bringing reported total funding to roughly $70M.

The current product line moves from scoring outputs to generating environments. Digital World Models are interactive simulations of real software: a ticketing system, a brokerage interface, a documentation site, a support console. An agent operates inside one of them and the simulation responds as the real product would, which lets a team observe failure modes, collect trajectories, and train on them without a live account or a real customer. Patronus says it has produced more than a million world data artifacts and that its simulated interfaces reach roughly 85 percent feature parity with the products they model. The buyers are frontier labs and enterprises that need agent behavior tested before it reaches a production system.