RL Environment Startups to Watch in 2026
Patronus AI TeamNovember 202612 min read
Patronus AI publishes this list and is one of the companies on it.
RL environment startups build the simulated worlds AI agents train in. In 2026 they fall into five groups. Research labs that build environments, such as Patronus AI, Andon Labs, Bespoke Labs and General Reasoning. Builders of software replicas and simulated workplaces, such as Fleet AI, Plato and Veris AI. Domain specialists such as Halluminate in finance and Mechanize in coding. Open hubs and platforms such as HUD and Prime Intellect. And game environments such as Good Start Labs.
This guide maps those startups, unranked and grouped by what they build. Each entry gives one dated fact, one piece of public work, the people behind it and whether it is hiring, so a buyer or a candidate can judge the work before talking to the team. For a ranked comparison that includes the larger data companies, see our guide to the best RL environment companies.
Last updated
What an RL environment is
An RL environment is a sandbox where an agent acts and gets scored. It has three parts. A task space, such as a codebase, a browser or a simulated app. A problem to solve, often written by people who do the work. And a verifier that decides whether the agent succeeded.
The verifier is where most of the difficulty lives. A task that is too easy teaches nothing, and a grader with a loophole teaches the model to exploit it, which is called reward hacking. Our guide to RL environments covers how they train an agent and how teams measure what they teach.
Five kinds of RL environment startups
The startups below build different things. Some publish research and build environments from it, some clone software and whole workplaces, some go deep in one domain, some give teams the tools to build their own, and one builds from games.
Research labs that build environments
Patronus AI, Andon Labs, Bespoke Labs and General Reasoning. They publish benchmarks and papers and build environments from that research.
Builders of software replicas and simulated workplaces
Chakra Labs, Fleet AI, Idler, Plato and Veris AI. They clone enterprise apps, websites or a company’s whole stack for agents to work in.
Domain specialists
Halluminate and Mechanize. They go deep in one field, finance or software engineering.
Open hubs and platforms
HUD and Prime Intellect. They give teams the tools and shared environments to build their own.
Game environments
Good Start Labs. It turns games into environments with verifiable rewards.
1. Research labs that build environments
These companies run research programs and turn the results into environments. Each publishes benchmarks or papers that anyone can read, so the work can be judged before a first call.
Patronus AI
Patronus AI is a research-first RL environment company. Its team publishes benchmarks like FinanceBench, TRAIL and GlobeBench alongside the environments it builds for AI labs.
- What it builds. RL environments and the data inside them for coding, computer use, knowledge work, science, alignment and more. Simulated apps, mock websites and the data behind them are also sold on their own.
- Latest release. On Nov 5, 2026 it announced GlobeBench, 1,632 tasks that test whether a model can predict how an environment responds to an agent’s next action. The best of 20 models solved 49.57% of tasks, and 631 tasks went unsolved by every model. The paper was accepted to the NeurIPS 2026 IAEval workshop.
- Funding. A $50M Series B led by Greenfield Partners, announced Jun 25, 2026.
- Public work. FinanceBench (2023), 10,231 questions about public companies’ filings, with 350+ citations and 82,000+ Hugging Face downloads of its open sample as of November 2026. TRAIL (2025), on finding errors in agent traces, with 110+ citations. Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis at ICML 2026. 18 papers in all.
- People. Founded in 2023 by Anand Kannappan and Rebecca Qian. Its researchers and engineers come from Meta Superintelligence Labs, Amazon AGI, Apple and Google.
- Hiring. Research and engineering roles at /join.
Andon Labs
Andon Labs is an AI safety lab that builds long-horizon simulated environments and runs businesses with AI agents in the physical world to compare against them.
- What it builds. Vending-Bench 2, where a model runs a simulated vending machine business for a year from a $500 starting balance. A full run takes 3,000 to 6,000 messages. Its other evaluations include Blueprint-Bench and Drone-Bench.
- Latest release. Drone-Bench, a paper on the drone surveillance capabilities of frontier models, on Jul 23, 2026.
- Dated event. On Apr 10, 2026 it gave an AI agent a three-year retail lease in San Francisco, the store now called Andon Market, and on May 4, 2026 it opened an AI-run café in Stockholm.
- People. Co-founder and CEO Lukas Petersson.
- Hiring. Researchers and engineers, through “Join the Lab” on its site.
Bespoke Labs
Bespoke Labs calls itself an applied AI research lab building RL environments and infrastructure so frontier labs and enterprises can train long-horizon agents.
- What it builds. Company-scale simulations with real codebases and microservices, a sandboxing layer that runs them, and agent optimization with GEPA and RL.
- Funding. $40M announced Jul 6, 2026, a $31.75M Series A led by Wing VC on top of an earlier $8.25M seed.
- Public work. OpenThoughts, an open reasoning dataset of more than a million examples. It is a core contributor to Terminal-Bench, the agentic coding benchmark.
- People. CEO Mahesh Sathiamoorthy, formerly of Google DeepMind, and CSO Alex Dimakis, a professor at UC Berkeley.
- Hiring. “Always hiring” researchers and engineers.
General Reasoning
General Reasoning is an AI research company that trains models to work over long horizons. Environment design is one of its research areas.
- What it builds. OpenReward, an open platform that serves 380+ community RL environments as API endpoints, built on the open Open Reward Standard.
- Latest release. Avalon-M1 on Oct 23, 2026, a model series that runs teams of agents rewarded as one.
- Public work. KellyBench (Apr 9, 2026), a simulated betting season of the 2023 to 24 English Premier League. Every model it evaluated lost money on average across five seeds.
- People. CEO Ross Taylor. The team previously drove open language models at Meta, including Galactica, Llama 2 and Llama 3.
- Hiring. Open roles on its careers page.
2. Builders of software replicas and simulated workplaces
These startups rebuild the software people use at work, from single apps to whole workplaces, so agents can practice office tasks with nothing real at stake.
Chakra Labs
Chakra Labs is a research and design lab that builds “task factories for useful intelligence”, expert-written tasks and pixel-perfect environments.
- What it builds. Tasks and environments for finance, legal, safety and computer use, including cloned enterprise apps.
- Latest release. MAGI, an MCP and GUI benchmark, on Oct 28, 2026. It holds 1,000 tasks, 200 for testing and 800 for training, across 8 cloned enterprise applications such as Slack, Linear, Notion and Grafana.
- Hiring. Three open roles, among them Member of Technical Staff in applied research.
Fleet AI
Fleet AI creates simulated worlds and challenges to understand and shape the behavior of AI, training gyms where agents practice tasks.
- What it builds. Simulated environments where agents practice workplace tasks. Its open source work is on GitHub.
- Funding. Runtime Wire reported a $45M Series A at a $725M valuation on Jun 1, 2026, led by Sequoia, Bain, Menlo and SVA. Fleet AI’s site does not mention it.
- Hiring. 13 roles in San Francisco and New York, among them environment generation and post-training research.
Idler
Idler builds RL environments from real production work.
- What it builds. Environments in three focus areas, frontier coding, economically valuable workflows and bespoke evaluations.
- Funding. Dealroom reported a $9M seed led by Paradigm on Aug 21, 2026, two days after the company launched, with Y Combinator and Long Journey Ventures also in.
- Hiring. Three open roles, among them Research Engineer and Offensive Security RL Researcher.
Plato
Plato builds simulated worlds for training and evaluating browser agents.
- What it builds. Simulated workplaces with the software and tasks of jobs in customer support, finance, accounting, HR and healthcare admin.
- Funding. A $16M Series A led by Nava Ventures, reported Oct 28, 2026, for $19M raised in all.
- People. Co-founder and CEO Robert Farlow.
- Hiring. No careers page on its site as of Nov 5, 2026.
Veris AI
Veris AI builds simulation infrastructure, a simulated copy of a team’s systems, data, APIs and users for agents to train and test in.
- What it builds. A stateful sandbox where coding agents verify their work. Scenarios, tool calls and user interactions are simulated at the tool and API boundary.
- Funding. An $8.5M seed led by Decibel and Acrew Capital, announced Jun 3, 2025.
- People. CEO Mehdi Jamei and CTO Andi Partovi, its co-founders.
- Hiring. Open positions linked from its homepage.
3. Domain specialists
These startups build for one field and hire people who know it, so their tasks and graders reflect how the work is done there.
Halluminate
Halluminate is a data research lab building benchmarks and RL environments for knowledge work, starting with financial services and consulting.
- What it builds. Computer use and tool use gyms on investment banking, private equity and consulting work, in desktop environments with trajectories that reach hundreds of steps.
- Funding. A $30M Series A led by Oak HC/FT on Oct 1, 2026, for $38.5M raised in all.
- Public work. Westworld Finance Diligence Bench (Aug 11, 2026), 88 problems that test an agent on a full company acquisition due diligence process.
- People. CEO Jerry Wu and CTO Wyatt Marshall. Founded in 2024.
- Hiring. Ten open roles, among them Member of Technical Staff in research and post-training, and Finance Researcher.
Mechanize
Mechanize builds environments and evals for frontier coding agents.
- What it builds. Software engineering tasks such as building a feature, deploying an application or debugging an issue in an unfamiliar codebase. A grader scores each run, and the scores serve as signals for RL and evaluation.
- Dated event. In September 2026, Business Insider reported that a talent and licensing deal with Google had closed. Co-founder Tamay Besiroglu joined Google DeepMind and Guive Assadi became CEO, and the license is reported as non-exclusive. The Next Web had reported talks worth more than $1.5 billion on Aug 6, 2026. Mechanize’s own site does not mention the deal.
- Funding. $9.1M at a $500M post-money valuation, announced Apr 24, 2026.
- Public work. GBA Eval, which measures whether coding agents can write a Game Boy Advance emulator from scratch in 24 hours, and essays such as “Cheap RL tasks will waste compute”.
- People. Co-founders Matthew Barnett, Tamay Besiroglu and Ege Erdil.
- Hiring. Software engineers, research engineers in alignment and a recruiter, in person in San Francisco.
4. Open hubs and platforms
These companies sell the tools rather than finished environments. Teams bring their own experts and build on a shared platform or start from community environments.
HUD
HUD is a platform for building RL environments, where teams encode their expertise into environments to train and evaluate models.
- What it builds. An SDK and community environments for coding, browser, computer use, deep research, robotics and more, with QA agents that audit traces for grader mistakes and reward hacking.
- Hiring. 37 open roles, 14 of them in engineering and research, such as Senior Software Engineer, Reinforcement Learning.
Prime Intellect
Prime Intellect builds an open stack for training models, with an Environments Hub of 2,500+ community RL environments.
- What it builds. The Environments Hub, built on the open source Verifiers library, plus hosted training on its prime-rl framework.
- Funding. $130M led by Radical Ventures, announced Jul 8, 2026.
- Public work. INTELLECT-3 (Nov 26, 2025), a 106B parameter mixture-of-experts model trained with SFT and RL on its stack.
- Hiring. 37 open roles, such as Research Engineer, RL Infrastructure.
5. Game environments
Games are compact worlds with clear rules and scores, which makes them natural RL environments for negotiation, strategy and humor.
Good Start Labs
Good Start Labs builds RL environments from curated games with verifiable rewards.
- What it builds. Game environments such as Diplomacy, where models negotiate and decide when to keep their word, and Bad Cards, where they pick the funniest answer. On Oct 1, 2026 it had models write StarCraft bots and climb a ladder of human-written ones.
- Funding. Spun out of Every with $3.6M from General Catalyst and Inovia Capital, Oct 15, 2025.
- Public work. Open Diplomacy and humor leaderboards, with 200,000+ matches played.
- People. Co-founder Alex Duffy, formerly head of AI training at Every.
- Hiring. Research scientists and machine learning engineers, remote in the US and Canada.
Larger companies that also sell environments
Expert data companies such as Scale AI, Surge AI, Mercor and Turing also sell RL environments, next to their human data and evaluation work. They are not startups in the sense of this list, so our ranked guide to the best RL environment companies covers them, and our guide to Scale AI alternatives compares the rest of the field.
They are also buyers. On Jul 9, 2026 Mercor announced it would acquire Deeptune, a startup building RL environments.
Comparison table
All fourteen startups at a glance, in the order of this guide. Latest dated news is the most recent dated release, round or deal we found on Nov 5, 2026.
| Company | Group | What it builds | Latest dated news | Public work |
|---|---|---|---|---|
| Patronus AI | Research lab | RL environments and their parts in hard domains | GlobeBench, Nov 5, 2026 | FinanceBench, TRAIL, TRACE, 18 papers |
| Andon Labs | Research lab | Long-horizon simulations and AI-run businesses | Drone-Bench, Jul 23, 2026 | Vending-Bench 2, Blueprint-Bench |
| Bespoke Labs | Research lab | Company-scale simulations and sandboxing | $40M, Jul 6, 2026 | OpenThoughts, Terminal-Bench |
| General Reasoning | Research lab | An open hub of community environments | Avalon-M1, Oct 23, 2026 | OpenReward, KellyBench |
| Chakra Labs | Software replicas | Cloned enterprise apps and expert tasks | MAGI, Oct 28, 2026 | MAGI benchmark |
| Fleet AI | Software replicas | Simulated workplaces for agents | $45M Series A, reported Jun 1, 2026 | Open source on GitHub |
| Idler | Software replicas | Environments built from production work | $9M seed, reported Aug 21, 2026 | Not listed on its site |
| Plato | Software replicas | Simulated workplaces for browser agents | $16M Series A, Oct 28, 2026 | Not listed on its site |
| Veris AI | Software replicas | Simulated copies of a production stack | $8.5M seed, Jun 3, 2025 | Not listed on its site |
| Halluminate | Domain specialist | Finance and consulting gyms | $30M Series A, Oct 1, 2026 | Westworld Finance Diligence Bench |
| Mechanize | Domain specialist | Coding agent environments and evals | Google talent and licensing deal, reported September 2026 | GBA Eval, essays on RL tasks |
| HUD | Hub or platform | A platform and SDK to build environments in-house | Not listed on its site | Community environments, SDK |
| Prime Intellect | Hub or platform | An open hub and training stack | $130M, Jul 8, 2026 | Verifiers, prime-rl, INTELLECT-3 |
| Good Start Labs | Games | Game environments | StarCraft bot ladder, Oct 1, 2026 | Diplomacy and humor leaderboards |
Where the market is heading
Labs are spending more on environments. Epoch AI’s Jan 12, 2026 review of the field cites a report that Anthropic discussed spending over $1 billion on RL environments over the following year, and notes a growing number of startups built only for this work.
Capital is following. Since June 2026, Bespoke Labs, Prime Intellect, Halluminate and Plato have each announced rounds, and Fleet AI and Idler were reported to have raised.
The market is also consolidating. Mercor agreed to acquire Deeptune in July, and Google’s talent and licensing deal with Mechanize was reported to have closed in September. Larger companies want environment teams in-house, so expect more deals like these.
How to choose an RL environment startup
Whether you are buying environments or looking for a job, the same few questions sort these companies.
Ask what ships
One startup’s environment is a full software replica with expert-written tasks and graders, another’s is a sandbox with a prompt. Ask for a sample and see what is inside.
Ask how graders are tested
A grader with a loophole teaches the model to cheat. Ask how the startup checks its verifiers against reward hacking, as the paper Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis does for code RL.
Read the public work
Benchmarks, papers and open source show how a team thinks before you meet it. Every entry above names at least one where we found it.
Check the stage
Funding dates and deals show who is growing and who may change hands. The table above lists the latest dated news for each.
For candidates, ask what you would build
Ask which domains you would work on, who writes the tasks and whether the team publishes. Each entry above links to where it is hiring.
Where Patronus fits
Patronus AI sits in the first group, research labs that build environments. It is a research-first RL environment company whose team publishes benchmarks like FinanceBench, TRAIL and GlobeBench alongside the environments it builds for AI labs.
Most of its work is custom. It builds environments in hard domains with AI labs, domain experts review every environment before delivery, and its research shapes the graders, for example the ICML 2026 paper Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis. It lists SOC 2, HIPAA and TISAX.
For a ranked comparison, see our guide to the best RL environment companies. To work on this research, see open roles.
Frequently asked questions
What is an RL environment startup?
A young company that builds the environments AI agents train in, or the tools to build them. Each environment gives an agent a task, the software to do it in and a grader that scores the result. Startups such as Patronus AI, Fleet AI, Mechanize and Prime Intellect sell these to AI labs and enterprises.
Which RL environment startups should I watch in 2026?
This guide covers fourteen in five groups. Research labs such as Patronus AI, Andon Labs, Bespoke Labs and General Reasoning. Software replica builders such as Chakra Labs, Fleet AI, Idler, Plato and Veris AI. Domain specialists Halluminate and Mechanize. Open hubs and platforms HUD and Prime Intellect. And Good Start Labs for games.
Which RL environment startups are hiring?
Most of them, as of Nov 5, 2026. Patronus AI, Fleet AI, HUD, Prime Intellect, Halluminate, Mechanize, Bespoke Labs, Good Start Labs and others list open research and engineering roles. Each entry above links to where it is hiring, and Patronus AI’s roles are at /join.
Which RL environment startups publish research?
Patronus AI publishes benchmarks such as FinanceBench, TRAIL and GlobeBench and 18 papers in all. Andon Labs publishes Vending-Bench 2, Bespoke Labs OpenThoughts and work on Terminal-Bench, and General Reasoning KellyBench. Chakra Labs, Halluminate and Mechanize publish benchmarks too.
How are RL environment startups different from data companies like Scale AI?
Startups build environments as their main product. Expert data companies such as Scale AI, Surge AI, Mercor and Turing sell environments next to large human data programs. Our ranked guide to the best RL environment companies compares both.
Are RL environment startups being acquired?
Some are. Mercor announced on Jul 9, 2026 that it would acquire Deeptune, and in September 2026 Business Insider reported that Google’s talent and licensing deal with Mechanize had closed.
Last updated