Patronus AI Raises $50M to Build Digital Worlds That Stress-Test AI Agents Before Deployment

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

AI agents are becoming more sophisticated. They are evolving from answering questions to autonomously executing multi-step complex tasks. But before these agents can be trusted to book trips or conduct financial analysis on behalf of users, model providers and the startups building such agents want to ensure that they perform reliably across a vast range of scenarios. AI labs often use benchmarks to show off their model's prowess, but a high score, even on an agent-oriented benchmark, doesn't actually prove that an AI can accomplish various complex, real-world jobs correctly. Patronus AI, a startup founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian, is helping model makers and companies fine-tune models to do just that by building simulated digital environments in which to evaluate the agents' performance. The San Francisco-based startup must be solving an important problem. Virtually every frontier AI lab and many emerging startups are now customers, according to Glenn Solomon, a managing director at Notable Capital, who describes demand for the company's simulated environments as nearly insatiable. Patronus' revenue has grown 15-fold over the past year, fueling significant 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. The funding brings the company's total funding to $70 million. Patronus uses what it calls "digital world models" to create replicas of websites and internal systems. In these environments, agents are stress-tested after training using reinforcement learning, which iteratively rewards successful task completion and penalizes errors. AI labs see great value in these digital simulations because they give agents a chance to try different, sometimes unpredictable, scenarios. The company compares its approach to how Waymo trained autonomous cars by first building synthetic worlds to test vehicles against rare hazards, such as severe weather or a child running after a ball. The difference with AI agents is that they tend to take shortcuts, which means they fail to complete the task correctly. "Patronus is really good at spotting the hacks and making sure they are holding the models accountable," Solomon said. Patronus is currently providing its simulated digital worlds for software engineering and finance, but these are just the start, according to Kannappan. "Today we're very focused on the problems that are verifiable, so the problems that you can immediately check and verify, but there are a ton more areas that are very non-verifiable or very hard to verify," he said. Just because these processes are verifiable doesn't mean they are simple. "We want to be able to actually create the environment in which you can operate an agent that can run for 10 hours or 10 days or 10 weeks," Kannappan said. As for rivals, Patronus believes it is primarily competing against the internal teams AI labs have already built to evaluate agent behavior. While human-data firms like Mercor and Surge help model makers with reinforcement learning, Patronus operates differently by evaluating how agents behave without any human involvement.

Patronus AI grabs $50M in funding to stress-test AI agents in simulated environments

Patronus AI grabs $50M in funding to stress-test AI agents in simulated environments Fast-growing world model startup Patronus AI Inc. is priming itself for even more rapid growth after raising $50 million in Series B funding today. The round was led by Greenfield Partners and saw the participation of Lightspeed Venture Partners, Notable Capital, Datadog and Samsung Ventures, and brings the company's total amount raised to date to $70 million. Patronus AI was founded by former Meta Platforms Inc. artificial intelligence researchers Anand Kannappan and Rebecca Qian, who are on a mission to ensure that autonomous agents can be put to work reliably. They're building the infrastructure to enable comprehensive AI agent training, so that other researchers can enhance the performance and reliability of AI systems spanning applications from financial trading to healthcare diagnostics and drone automation. The startup has enjoyed strong growth over the last year as AI systems become more sophisticated and capable. These days, AI doesn't just answer people's questions, but autonomously executes complex, multistep tasks on their behalf, such as booking tables at restaurants, buying and selling stocks at predetermined prices and more. However, autonomy can be risky, and before any AI agent is trusted to conduct such activities, there's a need to ensure that it will do the job as expected, without causing any problems or getting things wrong. This is where Patronus AI comes in. AI developers use benchmarks to demonstrate their AI model's performance and capabilities, but even a chart-topping score on an agent-oriented benchmark doesn't really mean much. The problem is that working autonomously in the real world is a completely different ball game as there are so many external factors that can impact an agent's ability to correctly fulfill a task. Patronus AI's world models enable developers and researchers to build simulated digital environments that more accurately reflect real world conditions, enabling agents to be put through their paces in multiple different scenarios. According to Notable Capital Managing Director Glenn Solomon, they're extremely popular, used by virtually every major AI lab and dozens of startups. He said the company is seeing "insatiable" demand for its simulated environments, and has increased its revenue by 15-fold in the last year. With the Patronus AI's world models, developers can create full working replicas of websites and corporate applications, where AI agents can be stress-tested after training them with reinforcement learning - a technique that involves rewarding agents for successfully completing tasks and penalizing them for failure. Within these simulated environments, AI agents can be tested in a wide range of unpredictable scenarios to see how they deal with the unexpected. It's similar to how Waymo LLC built a simulation to teach its autonomous cars to avoid hazards such as a child running after a ball. Kannappan said these kinds of simulations are necessary, because benchmarks only provide static evaluations that show if a model can perform in a tightly controlled setting. "They do not tell you whether an agent can navigate ambiguity, recover from failure or operate reliably across long, unpredictable workflows," he said. "That requires environments where systems can practice, adopt and accumulate experience over time." For now, Patronus AI is mainly focused on building simulated worlds for finance and software engineering tasks, but Kannappan said its ambitions extend well beyond this. "We're very focused on problems that are verifiable, so the problems that you can immediately check and verify, but there are a ton more areas that are very non-verifiable or very hard to verify," he told TechCrunch in an interview. The opportunity is especially compelling because Patronus AI seems to be operating in a very uncrowded niche, with few obvious rivals that can match its agentic testing capabilities. Kannappan said the company's biggest competitors are the internal model evaluation teams built up by AI labs. Other world model developers, such as Google LLC and Decart AI Inc., are more focused on AI training than performance evaluations. "Patronus AI is tackling one of the most important infrastructure problems in AI," said Greenfield Partners' Itay Inbar. "The future of AI will depend on systems that can learn and operate reliably in complex environments, and simulations are becoming essential to making that possible."