Reinforcement learning environments went from research plumbing to a funded vendor category in about a year. The reason is agents: once labs started training models to complete long-horizon tasks in software, they needed realistic places for those agents to practice, with verifiable rewards to learn from. Building a convincing replica of Salesforce, a hospital records system, or a full coding workflow is real work, and a wave of companies now sells exactly that.
This guide maps who builds RL environments in 2026, how the vendors differ, and the part of the stack that environments do not replace: the real-world data underneath.
What an RL environment actually is
An RL environment is a controlled world where an agent acts and gets scored. In practice that means three components: a simulated task space, such as a replica of enterprise software, a codebase, a browser, or an operating system; a goal specification, often written by domain experts; and a verifier that decides whether the agent succeeded, which is where most of the engineering difficulty lives. Weak verifiers get gamed, and reward hacking is the failure mode everyone in this market is quietly fighting.
Three kinds of vendors sell them
The first group is the human data companies that added environments to existing RLHF and evaluation businesses. Scale AI sells simulated web apps, desktop VMs, and tool-based environments. Surge AI built an enterprise simulation suite. Mercor added environments alongside its expert networks, including through acquisitions. Our guide to Mercor alternatives covers that company's full trajectory.
Turing and Centific pitch environments as part of broader data services; we cover Centific's version in our guide to Centific alternatives. The pitch here is one vendor for data, evals, and environments together.
The second group is environment-native startups, founded for this market specifically. Mechanize builds high-fidelity coding environments for frontier labs. Fleet AI replicates enterprise software like CRMs and spreadsheets in high fidelity. HUD wraps real software as agent-callable tools in containers. Veris AI and Plato build simulated enterprise and web worlds. Bespoke Labs comes at it from the tooling side, with open-source libraries for curation and evaluation. These companies tend to be deeper on a specific environment type and faster-moving than the incumbents.
The third group is open ecosystems and infrastructure. Prime Intellect hosts thousands of community-built environments with an open verifier library. General Reasoning runs a community hub with a unified API. Underneath all of it sit sandbox and compute providers like Modal and E2B, which run the containers everyone else's environments execute in.
The landscape
| Company | Category | Known for |
|---|---|---|
| Scale AI | Data giant | Coding environments over real and private codebases |
| Surge AI | Environment-native | High-fidelity coding environments for frontier labs |
| Mercor | Environment-native | Enterprise software replicas |
| Centific | Environment-native | Real software wrapped as agent-callable tools |
| Turing | Environment-native | Enterprise simulation plus production runtime |
| Mechanize | Environment-native | Open tooling for curation and evals |
| Fleet AI | Open ecosystem | Community hub of thousands of environments |
| HUD | Infrastructure | Sandboxed compute the environments run on |
| Veris AI | ||
| Bespoke Labs | ||
| Prime Intellect | ||
| Modal, E2B |
How to buy in a market this young
Three cautions. First, definitions vary wildly: one vendor's "environment" is a full software replica with expert-written goals and verifiers, another's is a sandbox with a prompt. Ask exactly what ships. Second, verifier quality is the product. A beautiful simulation with a gameable reward teaches your agent to cheat; ask how verifiers are tested against reward hacking. Third, most environment work is still custom, so treat public offerings as starting points and scope carefully, the same discipline we recommend for choosing any data vendor.
Environments do not replace data
Here is the part of the stack the environment wave tends to obscure. An environment teaches an agent to act, but the model inside the agent still has to understand the world it is acting in, and that understanding comes from training data. World models learn physics and object behavior from real video. Agents that operate software learn interfaces from real demonstrations. Game-playing and embodied models learn from gameplay data, task demonstrations, and first-person footage of humans actually doing things. The vendors supplying that physical side are mapped in our guide to robotics training data companies.
Simulation and real-world data are complements: the environment provides practice and verification, the data provides the ground truth the simulation is built from and the model is grounded in. Labs that skip the second half get agents that ace the benchmark and fumble reality.
Where Troveo fits
Troveo does not build RL environments. Troveo licenses the real-world interaction data that agent and world-model teams train on: gameplay footage, task demonstrations, tool manipulation, first-person navigation, and real workflows, sourced from more than 7,000 rights holders and cleared for AI training with documentation per asset. If you are standing up an environment program, that data is what grounds the model before the environment sharpens it. Browse the catalog in Lens or talk to us about what your agents need to learn from.