Patronus AI

Patronus AI is an agent-simulation and evaluation platform that builds interactive 'digital world models' — synthetic replicas of websites and internal systems — where enterprises reinforcement-learn and stress-test AI agents against rare failures before production. It is built for AI labs and enterprise teams deploying autonomous agents in software engineering, finance, and customer service.

Founded in 2023 by former Meta AI researchers Anand Kannappan and Rebecca Qian, San Francisco-based Patronus AI positions itself as a frontier lab for simulation research and infrastructure for AI agents. Its core product, the Digital World Model (DWM), dynamically generates interactive synthetic environments that mirror real products and workflows, so agents can be trained and stress-tested after the fact using reinforcement learning that rewards successful task completion and penalizes errors — an approach the company likens to how Waymo trained self-driving cars in synthetic worlds before the road. Patronus reports its worlds deliver 30-40% model lift on long-horizon tasks with over 1M world data artifacts and ~85% UI/UX feature parity with real products, and it also ships evaluation models and benchmarks including Lynx (hallucination detection), GLIDER (explainable LLM-as-judge), Percival, and FinanceBench. In June 2026 the company raised a $50M Series B led by Greenfield Partners, bringing total funding to $70M, and says revenue grew 15-fold over the prior year with adoption across virtually every major frontier AI lab.

Category

Governance & Security

Pricing

Contact for pricing

Target Market

CTOs, Heads of AI, ML Engineers, AI Researchers

Founded

2023

Headquarters

San Francisco, United States

Customers

Used by virtually every major frontier AI lab and dozens of startups

• ✓
  Digital World Models

  Dynamically generated interactive synthetic replicas of websites and internal systems where agents are trained and stress-tested.

• ✓
  Reinforcement-learning stress testing

  Iteratively rewards successful task completion and penalizes errors to harden agents against rare failures before deployment.

• ✓
  Lynx hallucination detection

  An evaluation model built to detect LLM hallucinations, which Patronus reports beats GPT-4 on hallucination tasks.

• ✓
  GLIDER evaluation model

  An explainable LLM-as-a-judge model that produces reasoning chains and highlights for scoring agent and model outputs.

• ✓
  Domain benchmarks

  Purpose-built benchmarks such as FinanceBench (10,000 finance Q&A pairs) for evaluating LLM and agent performance.

• •
  Pre-deployment agent testing

  Run agents through simulated digital worlds to surface failures before they reach real users or systems.

• •
  AI lab evaluation infrastructure

  Provide frontier labs with automated, human-free evaluation of agent behavior at scale.

• •
  Finance and software agent hardening

  Stress-test agents on verifiable finance and software-engineering tasks where errors are costly.

Public pricing is not listed; the company offers an interactive playground and documentation, with enterprise engagements via sales.

This page was written from 2 sources, 1 on domains other than patronus.ai.

1. 1.patronus.ai — patronus.aivendor
2. 2.techcrunch.com — patronus ai lands 50m to build digital worlds that stress te