Patronus AI raises $50M Series B to build Digital World Models that stress-test AI agents

Patronus AI has raised a $50 million Series B led by Greenfield Partners, alongside a preview of what it calls its first Digital World Model — software meant to simulate environments and steer how AI agents behave inside them. The round brings total funding to $70 million, with participation from Lightspeed Venture Partners, Notable Capital, Datadog, Samsung, Factorial Capital and investor Gokul Rajaram.

The pitch addresses a real gap: as companies hand more work to autonomous agents, there is no reliable way to predict how an agent will act before it is loosed on live systems. Patronus AI, which started as an AI evaluation company, says its Digital World Model uses language diffusion to predict environment behavior and steer agent actions across digital workflows — in effect, a simulator for stress-testing agents.

The company says it evaluated the model across coding benchmarks including InterCode and SWE-smith, the tau-bench dialogue benchmark, research and GUI tasks, and tool-use suites such as API-Bank and BFCL-v4. Those results are company-reported and not independently verified. Patronus AI also said its revenue has grown more than 15-fold in the past year and that it works with most of the world’s leading frontier AI labs and hyperscalers — figures it disclosed without external confirmation.

What the preview leaves open is whether a simulated world predicts real-world agent behavior well enough to be trusted as a safety check, the central claim the product rests on. The company has not released the model’s error rates or independent evaluations.

Patronus AI said the new money will expand its research and engineering teams and the compute behind the Digital World Models. The test will be whether enterprises adopt simulation as a standard gate before deploying agents, rather than testing in production.