Patronus AI lands $50M to build ‘digital worlds’ that stress-test AI agents

AI agents are becoming more sophisticated and capable of handling complex multi-step tasks, but before they can be trusted to perform real-world jobs like booking trips or conducting financial analysis, they need to be thoroughly tested. Patronus AI has built simulated digital environments, called "digital world models," that replicate websites and internal systems to stress-test these agents and ensure they work reliably across various scenarios. The startup creates replicas where agents are trained using reinforcement learning, which rewards successful task completion and penalizes errors, with the goal of identifying when agents take shortcuts rather than completing tasks correctly. The company announced a $50 million Series B funding round, bringing its total funding to $70 million, as virtually every major AI lab and many startups have become customers seeking to verify their agents' performance.