Key facts

Founded
2023
HQ
San Francisco, United States
Employees
—
Funding stage
Series B · US$50m
Further facts
  • Funding: In June 2026 Patronus AI raised a US$50m Series B led by Greenfield Partners, with Lightspeed Venture Partners, Notable Capital, Datadog and Samsung taking part.
  • Research: Patronus has released agent benchmarks including TRAIL for finding where an agent went wrong, MEMTRACK for agent memory and SpeedrunBench.

Focus areas

LLM evaluation AI observability Agentic AI AI governance Physical AI

About

Patronus AI is a San Francisco behavioural research lab for AI, founded in 2023 by Anand Kannappan and Rebecca Qian. Its researchers and engineers come from Meta, Amazon, Apple and Google. Patronus started with tools to evaluate and test large language models and AI agents for reliability, and now also builds simulated environments, post-training data and benchmarks that AI labs and enterprises use to train and measure agents across coding, research, knowledge work and computer use. Lightspeed Venture Partners led its seed round and Notable Capital its Series A. In June 2026 it raised a US$50m Series B led by Greenfield Partners and released a preview of its first Digital World Model; the money funds research, sales and the compute to run these models at scale.

How they use AI

Patronus builds automated judges, benchmarks and simulation environments that show where models and agents fail. Its Digital World Models are language diffusion models that predict how software environments behave, so an agent can practise long, multi-step tasks in simulated workplaces instead of fixed datasets. In its own tests the model scored highest on coding, dialogue, research and general tool use against two frontier models. Generative Simulators create new training environments automatically, and public benchmarks such as TRAIL and MEMTRACK measure agent error tracing and memory across workplace tools.

Products and services

  • Digital World Models — Language diffusion models that simulate software environments for agent training
  • Generative Simulators — Automatically scaled training environments for agents
  • Post-training data — Datasets and environments for coding, research, knowledge work and computer use
  • LLM testing and AI reliability — Evaluation of model and agent outputs for errors and failures

Key people

  • Anand Kannappan — Co-founder
  • Rebecca Qian — Co-founder

Investors