Patronus AI raises $50M to build digital worlds for testing agents

Patronus AI announced a $50 million Series B round led by Greenfield Partners, bringing the company’s total funding to $70 million. The San Francisco startup, founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian, builds simulated "digital world models" to evaluate how AI agents behave in replicated websites and internal systems.

What does Patronus AI build and how does it work?

Patronus builds "digital world models" that create replicas of websites and internal systems, then runs trained agents inside those simulations and evaluates their behavior using reinforcement learning. In those environments the company rewards successful task completion and penalizes errors, allowing agents to encounter unpredictable scenarios and revealing when models take shortcuts or fail to finish tasks correctly.

The startup compares its approach to how Waymo used synthetic worlds to test autonomous vehicles against rare hazards. Patronus emphasizes hands-off evaluation: it measures agent behavior without human involvement, rather than relying on human-data firms to guide reinforcement learning. The company currently focuses on software engineering and finance use cases, with the idea that simulations should allow agents to run for extended periods. As Anand Kannappan put it, "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."

Who backed the round and how fast is the business growing?

The $50 million Series B was led by Greenfield Partners, with participation from Notable Capital, Lightspeed, Datadog, and Samsung. That round brings Patronus’ total financing to $70 million. Glenn Solomon, a managing director at Notable Capital, described demand for Patronus’ simulated environments as "nearly insatiable."

Patronus’ revenue has grown 15-fold over the past year, a metric investors cited as a signal of traction. The company was founded in 2023 by Anand Kannappan and Rebecca Qian, both former Meta AI researchers.

Why does this matter?

AI labs and startups are building agents that must perform multi-step, real-world tasks, and benchmark scores do not always show whether an agent can complete those tasks reliably. Patronus’ simulations expose shortcuts and rare failure modes by reproducing the systems agents will face in production. The result is a more rigorous, replicable evaluation pipeline that model providers can use to fine-tune agents before deployment.

Patronus positions itself against internal evaluation teams at AI labs and against human-data firms such as Mercor and Surge, differentiating on automated, non-human evaluation. Glenn Solomon said the company is "really good at spotting the hacks and making sure they are holding the models accountable," a direct appeal to clients worried about brittle or gamed behaviors.

What to watch

Watch for Patronus expanding beyond its current software engineering and finance offerings into harder-to-verify domains, and for demonstrations of long-running agent evaluations at the durations Anand Kannappan described (10 hours, 10 days, 10 weeks). Another milestone will be whether major frontier labs standardize on Patronus’ simulated environments instead of building larger in-house evaluation suites.

Patronus’ next public signals will likely be new verticals, customer case studies showing prevented failures, or technical disclosures about how its digital worlds surface agent shortcuts and verifiable failure modes.