Patronus AI押注模拟世界而非基准测试来验证智能体；获5000万美元B轮融资

做的是什么生意

Simulation and evaluation infrastructure for AI: Patronus builds Digital World Models, large-scale replicas of websites, software and internal systems, where AI agents train, get stress-tested and improve before touching real systems.

启动资金：$50M Series B led by Greenfield Partners (June 2026) with Notable Capital, Lightspeed, Datadog and Samsung participating; brings total funding to $70M.

起因

Patronus AI was founded in 2023 in San Francisco by former Meta AI researchers Anand Kannappan and Rebecca Qian. Having worked on evaluation and alignment inside a frontier lab, they argued that static benchmarks tell you whether a model answers a narrow question, not whether an agent can navigate ambiguity, recover from failure and run long workflows reliably.

经过

Patronus grew from LLM evaluation tools into simulation infrastructure as AI agents moved from answering questions to executing multi-step work. By mid-2026 the company said it worked with the majority of the world's leading frontier AI labs and hyperscalers, and revenue grew more than 15x over the preceding year. On June 25, 2026 it announced a $50M Series B led by Greenfield Partners, with Notable Capital, Lightspeed, Datadog and Samsung, and unveiled Digital World Models, comparing the approach to Waymo testing self-driving cars against rare hazards in synthetic worlds before real roads.

结果

Scaling: with $70M raised, Patronus plans to expand its research and engineering organizations and invest in the compute needed to train and run Digital World Models at scale, starting with verifiable software-engineering and finance workflows and aiming toward agents that operate for days or weeks at a time.

可借鉴之处

智能体在基准测试无法衡量的冗长、模糊任务上会失败；通过销售用于真实工作流的模拟环境，评估基础设施成为前沿实验室必须购买的一层。