Patronus AI raises $50M to stress-test AI agents in simulated worlds

Patronus AI has closed a $50 million Series B round to expand its platform for evaluating AI agents in simulated digital environments, the company announced Thursday. Greenfield Partners led the round, with participation from Notable Capital, Lightspeed, Datadog, and Samsung, bringing total funding to $70 million.

The San Francisco startup creates what it calls "digital world models"—replicas of websites and internal systems where AI agents can be tested against complex, multi-step tasks before being deployed to real users. The approach mirrors how Waymo trained autonomous vehicles in synthetic environments before road testing, allowing agents to encounter rare or unpredictable scenarios safely.

Founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian, Patronus addresses a gap between benchmark performance and real-world reliability. While AI labs routinely publish high scores on agent-oriented benchmarks, those numbers don't prove an AI can correctly book travel, conduct financial analysis, or complete other autonomous tasks users might assign.

Testing agents through reinforcement learning

Patronus uses reinforcement learning in its simulated environments, iteratively rewarding successful task completion and penalizing errors. This training method helps identify when agents take shortcuts that technically complete a task but fail to do so correctly.

"Patronus is really good at spotting the hacks and making sure they are holding the models accountable," said Glenn Solomon, managing director at Notable Capital, who described demand for the company's simulated environments as nearly insatiable.

The platform currently focuses on verifiable domains like software engineering and finance, where outcomes can be immediately checked. But Kannappan said the company plans to expand into areas where verification is harder. The goal is to create environments where agents can operate for extended periods—"10 hours or 10 days or 10 weeks," he noted.

Why it matters

As AI agents evolve from answering questions to autonomously executing complex workflows, companies need reliable ways to validate their behavior before deployment. Traditional benchmarks measure narrow capabilities, but don't capture whether an agent will perform correctly across the messy, variable conditions of real-world use. Patronus's revenue grew 15-fold over the past year, suggesting model makers and enterprises see simulated testing as essential infrastructure for the agent era.

Customer base spans frontier labs

Virtually every major AI lab and many emerging startups now use Patronus, according to Solomon. The company positions itself as competing primarily against internal evaluation teams that AI labs have built in-house. While human-data firms like Mercor and Surge assist with reinforcement learning, Patronus differentiates by evaluating agent behavior without human involvement in the testing loop.

Details of the funding and customer traction were first reported by TechCrunch.