Patronus AI secures $50M to build digital worlds that stress-test AI agents

AI agents are growing increasingly sophisticated, shifting from answering questions to independently carrying out complex, multi-step tasks.

However, before these agents can be trusted to book travel or handle financial analysis on behalf of users, model providers and the startups developing them need to ensure reliable performance across a wide range of scenarios.

AI labs frequently rely on benchmarks to demonstrate their models' capabilities, yet a high score—even on an agent-specific benchmark—does not guarantee that an AI can correctly complete diverse, real-world tasks.

Patronus AI, a 2023 startup founded by ex-Meta AI researchers Anand Kannappan and Rebecca Qian, assists model makers and companies in fine-tuning models to achieve this by creating simulated digital environments that evaluate agent performance.

This San Francisco-based startup is clearly addressing a critical need. According to Glenn Solomon, a managing director at Notable Capital, virtually every frontier AI lab and many emerging startups have become customers, with demand for the company's simulated environments described as nearly insatiable.

Patronus's revenue has surged 15-fold over the past year, sparking strong investor interest. On Thursday, the company announced a $50 million Series B round led by Greenfield Partners, with participation from Notable Capital, Lightspeed, Datadog, and Samsung. This brings total funding to $70 million.

Patronus employs what it calls "digital world models" to build replicas of websites and internal systems. In these environments, agents undergo stress testing after training via reinforcement learning, which iteratively rewards successful task completion and penalizes mistakes.

AI labs find these digital simulations highly valuable because they allow agents to experiment with various, sometimes unpredictable, scenarios. The company likens its approach to how Waymo trained autonomous cars—by first constructing synthetic worlds to test vehicles against rare hazards like severe weather or a child chasing a ball.

The difference with AI agents is that they often take shortcuts, leading to incorrect task completion. "Patronus is really good at spotting these hacks and ensuring the models are held accountable," Solomon said.

Patronus currently offers its simulated digital worlds for software engineering and finance, but according to Kannappan, these are only the beginning.

"Today we're very focused on verifiable problems—those that can be immediately checked and verified—but there are many more areas that are non-verifiable or extremely difficult to verify," he said.

Just because these processes are verifiable doesn't mean they are simple. "We want to be able to create an environment where you can operate an agent that can run for 10 hours, 10 days, or even 10 weeks," Kannappan said.

As for competitors, Patronus believes it mainly competes against the internal teams that AI labs have already established to evaluate agent behavior. While human-data firms like Mercor and Surge assist model makers with reinforcement learning, Patronus takes a different approach by evaluating agent behavior without any human involvement.