RL Environment Companies Dataset, November 2026
Patronus AI TeamAs of November 20266 min read
A dated snapshot. Patronus AI publishes this dataset and is one of the companies in it. Companies are listed alphabetically, not ranked. Each value comes from the company’s own pages, announcements, GitHub, papers, LinkedIn page, Y Combinator profile or PR Newswire release, and links to it. Checked Nov 11, 2026.
This dataset lists 23 companies that build RL environments, expert data for AI training or the tools to run them, as of November 2026. For each one it gives the homepage, the year founded, headquarters, a one-line focus and its named public environments or benchmarks, with a link to the primary source of every value. Download it as CSV or JSON.
The set is the companies covered in our RL environment companies and RL environment startups guides. It records only facts that rarely change, so the snapshot stays useful without constant updates. Where a company does not state a fact on its own pages, the cell says so rather than guessing. The same companies appear by kind on our market map.
Snapshot as of November 2026. Last updated
Download the dataset
Both files hold the same 23 rows as the table below, plus a source URL for every field and the date each row was checked.
The dataset
23 companies, alphabetical. The small number after each value opens its source. Funding, headcount, open roles and prices are left out, because they change often.
| Company | Founded | HQ | Focus | Public work | As of |
|---|---|---|---|---|---|
| Andon Labs | 2023 1 | San Francisco, California 2 | Andon Labs studies frontier AI safety by running AI-run businesses in the real world and comparing models in simulated long-horizon evals. 3 | Vending-Bench 2, Drone-Bench (Jul 23, 2026 date), Butter-Bench (Oct 23, 2025) arXiv v1 date. Andon’s publications page lists Oct 28, 2025. | |
| Bespoke Labs | Not stated by the company | Mountain View, California 2 | Bespoke Labs builds RL environments and infrastructure so frontier labs and enterprises can train long-horizon agents for production. 3 | OpenThoughts, Curator, AutoResearchExam (Sep 10, 2026) | |
| Browserbase | 2024 1 | San Francisco, California 2 | Browserbase runs cloud browsers for AI agents, and its BrowserEnv wraps those sessions as an environment to train and evaluate browser agents. 3 | Stagehand, BrowserEnv, Browserbase MCP server | |
| Chakra Labs | Not stated by the company | Offices in New York, New York 2 | Chakra Labs is a research and design lab building expert-authored tasks and pixel-perfect environments to train models and evaluate real-world capability. 3 | MAGI (Jul 20, 2026 date), GDPx legal domain tasks (SemiAnalysis) (Nov 1, 2026 date) | |
| E2B | 2023 1 | Offices in San Francisco, California and Prague, Czech Republic 2 | E2B runs isolated microVM sandboxes for AI agents, in its cloud or the customer's, for coding, computer use, background jobs and RL rollouts. 3 | E2B, E2B Desktop, E2B runtime | |
| Fleet | Not stated by the company | San Francisco, California 2 Its LinkedIn lists New York. | Fleet builds training gyms for agents, high-fidelity simulation environments where AI practices tasks and humans provide supervision. 3 | WitnessBench (Sep 26, 2026) | |
| General Reasoning | 2025 1 | London, United Kingdom 2 | An AI research company training models for long-horizon work, with environment design as a research area and OpenReward, an open platform serving RL environments. 3 | OpenReward, Open Reward Standard (ORS), KellyBench (Apr 9, 2026) | |
| Good Start Labs | 2025 1 | Not stated by the company | Builds reinforcement learning environments and world-model data from games, plus public benchmarks that measure model capability in games. 3 | AI Diplomacy (Jun 5, 2025 date), Democratizing Diplomacy: A Harness for Evaluating Any Large Language Model on Full-Press Diplomacy (Aug 10, 2025), Humor leaderboard (Bad Cards) | |
| Halluminate | 2024 1 | San Francisco 2 | A data research lab building benchmarks and RL environments for knowledge work, starting with financial services and consulting. 3 | Westworld Finance Diligence Bench (Aug 11, 2026), Westworld v1 (Nov 23, 2025), Web Bench (May 30, 2025) | |
| HUD | 2025 1 | San Francisco, California 2 | A platform and SDK for building RL environments, turning expertise into environments that train and evaluate models, with QA agents auditing traces. 3 | hud-python SDK, SheetBench-50 (Oct 1, 2025), Assemble Bench (Jul 31, 2026) | |
| Idler | 2025 1 | San Francisco 2 | A frontier data lab building RL environments from real production work, across frontier coding, economically valuable workflows and bespoke evaluations. 3 4 | ShelfLife, ShelfLife E-Sim | |
| Invisible Technologies | 2015 1 | Not stated by the company | Builds expert-built RL environments where domain experts define task logic, design reward rubrics and annotate trajectories, alongside expert data. 3 | None found on its site | |
| Labelbox | 2018 1 | San Francisco, California 2 | Labelbox builds RL training environments and evals for post-training on reasoning, tool use and computer use, co-designed with domain experts. 3 | Implicit Intelligence (Nov 21, 2025), EchoChain (Mar 4, 2026), Labelbox Leaderboards | |
| Mechanize | Not stated by the company | Not stated by the company | Mechanize builds RL environments and evals in which frontier coding agents do software engineering tasks scored by a grader. 3 | GBA Eval (May 2, 2026), gba-eval (GitHub) | |
| Mercor | 2023 1 | San Francisco, California 2 | Mercor supplies expert human work for frontier research, AI benchmarks and AI agent training for AI labs and enterprises. 3 | APEX-Agents (Jan 20, 2026), APEX-SWE (Mar 24, 2026), APEX-Accounting (Jul 31, 2026) | |
| Patronus AI | 2023 1 | San Francisco 2 | Patronus AI is a behavioral research lab for AI that studies where models fail and builds simulated environments and training data to improve them. 3 | GlobeBench (Oct 15, 2026), TRAIL (May 13, 2025), FinanceBench (Nov 20, 2023) | |
| Plato | 2025 1 | San Francisco, California 2 | Plato builds simulated worlds of the software and tasks of real jobs, where models train and are evaluated on that work. 3 | Computer Work Val v1.0 (Oct 27, 2026), Web Clone Bench v1 (Oct 5, 2026) | |
| Prime Intellect | Not stated by the company | Not stated by the company | Prime Intellect offers an open-source library and hub for building and sharing RL environments, plus hosted RL training, evals and sandboxes. 3 | Environments Hub (Aug 27, 2025), Verifiers, INTELLECT-3 (Nov 26, 2025) | |
| Scale AI | 2016 1 | San Francisco, California 2 | Scale AI's Enterprise Atlas RL environments train agents on long-horizon work across real enterprise systems and desktop environments. 3 | SWE Atlas (Mar 4, 2026), AgentEnv (Oct 5, 2026) | |
| Snorkel AI | 2019 1 | Offices in San Francisco, California; New York, New York 2 | Snorkel AI develops specialized training data and environments for frontier labs and AI teams' models and agents. 3 | Agentic Coding 2.0, SnorkelFinance 2.0, SnorkelLegal | |
| Surge AI | 2020 1 | San Francisco, California 2 | Surge AI sells off-the-shelf training data, evals and RL environments built by the experts behind its frontier-lab work. 3 | DAYJOB: Tax, DAYJOB: Finance, RL Environments and the Hierarchy of Agentic Capabilities (Nov 3, 2025) | |
| Turing | 2018 1 | San Francisco, California 2 | Turing builds iterable UI and non-UI RL environments, including app replicas and MCP backends, for agents to learn long-horizon workflows. 3 | Turing Terminal-Bench Data, Turing Frontier Agentic STEM | |
| Veris AI | Not stated by the company | New York, New York 2 | Veris AI builds a simulated copy of a company's systems, data, APIs and users where AI agents are tested and trained. 3 | VAmoS Bench, VAmoS Pro Bench (Oct 1, 2026), Veris FinAgent Arena |
Fields and method
We started from every company named in our RL environment companies, RL environment startups and use-case roundups, and read what each company publishes itself on Nov 11, 2026. That means its company pages, announcements, GitHub and papers, and for 14 founding years and headquarters its LinkedIn page, its Y Combinator profile or its own release on PR Newswire. News articles and other company databases were used only to find those pages, never as the source of a value.
Founded is the year the company states. HQ is the city the company names as its headquarters, or the offices it lists when it names no headquarters. Focus closely follows the company’s own description of what it builds. Public work lists up to three environments, benchmarks, datasets or open source projects the company publishes under its own name.
- name
- Company name as it writes it
- url
- Company homepage
- group
- Kind of company, as on our market map
- founded
- Year founded, as the company states it
- hq
- Headquarters city, or listed offices when no headquarters is named
- focus
- One line on what it builds, after its own description
- public_work
- Named public environments, benchmarks, datasets or open source, with links and dates where stated
- *_source
- The primary source URL for each field. For public work, a second URL after a space is the page that states the date
- *_note
- A conflict between the company’s own sources, where one exists
- as_of
- The date the row was checked
Terms of use and corrections
Use of this dataset is covered by the Patronus AI Terms of Use.
To correct a value, send the company’s own source through our contact page.
Frequently asked questions
Is there a list of RL environment companies I can download?
Yes. This dataset lists 23 RL environment and expert data companies as of November 2026, as CSV and JSON. Each value links to the company’s own page it came from.
What fields does the dataset include?
Company name, homepage, kind of company, year founded, headquarters, a one-line focus and named public environments or benchmarks, with a source URL for every field and the date each row was checked.
Why does the dataset leave out funding and pricing?
Funding, headcount, open roles and prices change often, so a dated snapshot of them goes stale quickly. The dataset keeps facts that rarely change. Our RL environment startups guide covers recent rounds, dated.
How current is the dataset?
Every row was checked on Nov 11, 2026, and each row carries its own as-of date. It is a snapshot, so a newer company or change after that date is not in it.
Patronus AI · November 2026 snapshot · More resources