Patronus AI is an AI research company that builds training data and simulated environments for AI agents. It works with teams that train and post-train models, giving their agents realistic places to practise, along with expert-reviewed data and evaluations that show where the models fail and how to improve them.
Its environments and data span alignment, coding, knowledge work, science and computer use, and are modelled on real working environments. Alongside this commercial work the company publishes research papers, open models and benchmarks on agent behaviour, hallucination and AI judges.
At A Glance
| Vendor | Patronus AI |
|---|---|
| Category | RL environments, post-training data and evaluations for AI agents |
| Deployment | Custom builds delivered to the client’s own training harness |
| Agentic AI | Environments, trajectories and benchmarks built for AI agents |
| Certifications | SOC 2, HIPAA, TISAX |
| Funding | Seed, Series A and Series B rounds, latest in June 2026 |
| Market classification | AI Safety & Guardrails; also AI Model Security |
What Patronus AI Does
Patronus AI works as a behavioural research lab for AI. It builds environments where models can operate, studies how they behave there, and turns what it learns into data a model can be trained on. A model practising in a Patronus AI environment works through tasks, is scored by graders and uses simulated apps, so its developer can see the exact points where it goes wrong. Most engagements are custom: the company builds the tasks in each environment and calibrates them to the client’s needs. Its open research on how models fail feeds back into that work.
Products
RL Environments
RL environments are sandboxes with tools, a task and a reward that scores the result. They are for model developers who need somewhere realistic for an agent to learn, whether off the shelf or built to order, in any of the five domains. Parts can be bought on their own, such as simulated apps, mock websites and the data behind them, and every environment is reviewed by domain experts and checked automatically before delivery.
Post-Training Data
Post-training data comes out of the same custom builds, and ownership of the data is set in each agreement. Two kinds stand out:
- Expert trajectories for computer use, showing how a skilled person completes a task
- Expert pairwise reviews of agent transcripts, where a reviewer marks the better of two runs of the same task
Evaluations
Evaluations tell a client how well a model handles its tasks. Patronus AI combines expert manual review with automated checks, and also trains open models on its own tasks and tests them on held-out tasks.
Digital World Model
The digital world model simulates digital environments for AI agent training, so agents can be trained and tested in conditions that mimic real scenarios without a hand-built environment for every task.
Open Models And Benchmarks
| Lynx | An open-source model that detects hallucinations in retrieval-augmented generation |
|---|---|
| GLIDER | A small, explainable judge model for grading LLM interactions |
| FinanceBench | Open-book financial question answering about publicly traded companies |
| TRACE | 54 kinds of reward exploit in code RL, and how well models catch them |
| SimpleSafetyTests | 100 prompts for finding critical safety risks in language models |
How It Works
Delivery is services-led. A typical engagement starts with scoping, where Patronus AI confirms how the tasks will run on the client’s own training harness, and the company usually sends samples before the full build.
| Component | What it does |
|---|---|
| Environment | A sandbox with tools, a task and a reward, in which an agent runs and is scored. |
| Simulated apps and data | Mock software and websites, with their underlying data, that make the environment behave like a real workplace. |
| Review | Domain experts review each environment and automated checks run before it ships. |
| Expert network | Domain specialists from an expert network contribute to the data and reviews. |
Agentic AI Capabilities
Everything Patronus AI sells is aimed at AI agents. Its computer use environments are where agents learn to get work done in software built for people. Its research focuses on how agents fail:
| TRAIL | A benchmark for finding the step where an agent went wrong |
|---|---|
| MEMTRACK | Tests whether agents keep track of facts spread across workplace tools |
| DETOUR | A two-agent benchmark for tip-of-the-tongue search |
| World Modeling for Agentic RL | A language model that stands in for an environment while an agent trains |
| Generative Simulators | Environments that generate their own tasks, world dynamics and rewards |
Company Background
| Founded | 2023 |
|---|---|
| Leadership | Anand Kannappan (co-founder and CEO), Rebecca Qian (co-founder and CTO) |
| Funding | $3M seed in September 2023 led by Lightspeed Venture Partners; $17M Series A in May 2024 led by Notable Capital; $50M Series B in June 2026 led by Greenfield Partners |
| Investors | Greenfield Partners, Lightspeed Venture Partners, Notable Capital, Datadog, Samsung Ventures and InvestInData |
Note: This overview is compiled from Patronus AI’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.