Patronus AI just closed a $50 million Series B to build simulated digital worlds where AI agents get put through their paces before they’re trusted with real tasks. The round, announced Thursday and reported by TechCrunch AI, was led by Greenfield Partners, with Notable Capital, Lightspeed, Datadog, and Samsung joining in. That brings the San Francisco startup’s total funding to $70 million.
Patronus was founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian. Their pitch is simple: a high benchmark score doesn’t prove an agent can actually book your flight or run a financial analysis without messing it up. So they build the testing ground that does.
What Patronus actually builds
The company uses what it calls “digital world models” to create working replicas of websites and internal systems. Agents get dropped into these environments and stress-tested after training, using reinforcement learning that rewards finishing a task correctly and penalizes mistakes.
TechCrunch AI reports that Patronus compares its approach to how Waymo trained self-driving cars. Before putting vehicles on real roads, Waymo built synthetic worlds to test them against rare hazards like severe weather or a kid chasing a ball into the street. Patronus does the same thing for software agents.
There’s one twist specific to AI agents. They love shortcuts. An agent will often find a way to technically “complete” a task while skipping the part that actually matters.
“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.
Why this matters
AI agents are moving from answering questions to running multi-step jobs on their own. That shift only works if the agents are reliable, and right now the industry’s main proof point is benchmarks. Benchmarks are easy to game and don’t reflect messy real-world conditions. Patronus is selling the missing layer: a way to see how an agent behaves when things get unpredictable.
What stands out here is the demand signal. Solomon describes appetite for these simulated environments as nearly insatiable, and says virtually every frontier AI lab plus many emerging startups are now customers. Revenue grew 15-fold over the past year. When a two-year-old company posts that kind of curve, it usually means it’s solving a problem the whole market feels at once.
Where it goes next
Right now Patronus focuses on areas you can check and confirm, mainly software engineering and finance. Kannappan calls these “verifiable” problems, but he’s clear that verifiable doesn’t mean easy.
“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.
The bigger ambition is duration. Kannappan wants environments that can run an agent for far longer than a quick test. “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,” he said. Long-running agents are where a lot of the real value sits, and also where small errors compound into big failures.
The competition
Patronus says its main rival isn’t another vendor. It’s the internal evaluation teams that AI labs have already built in-house. Human-data firms like Mercor and Surge help model makers with reinforcement learning, but Patronus works differently. It evaluates how agents behave with no human in the loop at all.
That’s the bet worth watching. If agents are going to act on our behalf, someone has to test them at scale without a person checking every step. Patronus thinks that job is too big for labs to keep doing alone, and $50 million says a lot of investors agree.
For the full breakdown, check the original report at TechCrunch AI.