Patronus raises $50M to stress-test AI agents reliably

Patronus AI, founded by former Meta researchers in 2023, has raised a $50 million Series B to expand its simulated “digital world models.” The San Francisco startup says it helps AI labs and companies fine-tune agents so they can handle real-world tasks withou

By the time an AI agent is trusted with booking trips or running parts of an analysis, the hard part isn’t just intelligence—it’s reliability. Patronus AI is betting that reliability has to be tested in environments that look nothing like the polished benchmarks developers typically brag about.

The startup. based in San Francisco and founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian. is building simulated “digital world models” designed to evaluate whether an agent can actually finish complex. multi-step jobs. Instead of relying on a high score on an agent-oriented test. Patronus aims to push models through a range of real-world scenarios where failure can be costly.

Demand is climbing fast. Glenn Solomon, a managing director at Notable Capital, says the company’s simulated environments have nearly “insatiable” demand among virtually every frontier AI lab and many emerging startups.

On Thursday, Patronus announced a $50 million Series B round led by Greenfield Partners. Notable Capital, Lightspeed, Datadog, and Samsung also participated. The new funding brings Patronus’ total funding to $70 million. and it comes after the company’s revenue grew 15-fold over the past year—growth that has helped draw significant investor attention.

Patronus’ approach revolves around creating replicas of websites and internal systems. After models are trained using reinforcement learning, agents are stress-tested inside these environments. Successful task completion is rewarded; errors are penalized—an iterative loop meant to surface the places where agents can look capable while still failing when conditions change.

Solomon said Patronus is “really good at spotting the hacks and making sure they are holding the models accountable.” The reason matters: AI agents, he argues, often take shortcuts that allow them to pass narrow tests without actually completing the underlying task correctly.

The company frames its simulated-world strategy as similar in spirit to how Waymo trained autonomous cars by building synthetic environments to test against rare hazards. such as severe weather or a child running after a ball. The difference is what engineers are trying to prevent in software. With AI agents. the concern is not a collision—it’s the moment an agent “finds” a shortcut that defeats the intent of the job.

For now, Patronus is providing simulated digital worlds for software engineering and finance, but Kannappan says those are only the beginning. He emphasized that while some tasks are immediately measurable, many of the problems worth solving don’t fit neatly into verification.

“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.

That time horizon is part of the pitch: the environment has to support agents running for long stretches, not just completing a single prompt and moving on.

Patronus also sees a clear competitive gap. It believes it is primarily up against the internal teams AI labs already build to evaluate agent behavior. While human-data reinforcement learning firms like Mercor and Surge help model makers train systems. Patronus positions itself differently—evaluating how agents behave without any human involvement.

Patronus AI digital world models AI agents reinforcement learning simulated environments stress-testing Series B Greenfield Partners Notable Capital Lightspeed Datadog Samsung Anand Kannappan Rebecca Qian