Worlds agents train in
RL environments for AI agents: what each builder says it makes
Updated
Builders describe an environment by what the agent has to do inside it. This page takes five of the 12 payers on the board, Mechanize, Veris AI, Prime Intellect, Halluminate and Collinear, and groups their words by the kind of agent.
For coding agents
Mechanize's home page says it builds environments and evals for frontier coding agents, where models build a feature, deploy an application or debug an issue in an unfamiliar codebase.
Source: What is paid to build RL environments and UI gyms: every row, with its source
Veris AI's home page says it gives coding agents a stateful simulation of a customer's dependencies, so integration tests run and issues are fixed before the pull request opens, and Prime Intellect's Environments Hub lists DeepSWE long-horizon software engineering tasks from Harbor Hub.
Source: What is paid to build RL environments and UI gyms: every row, with its source
For computer-use agents
Halluminate says it builds realistic sandboxes and datasets to train computer and browser use AI, in dynamic desktop environments with trajectories that reach hundreds of steps. Collinear's home page describes each simulation lab as a self-contained world with the users, tools, data and tasks an agent will meet in production, from GUIs and APIs to MCP, CLI, GitHub, Jira, ServiceNow and Salesforce.
Source: What is paid to build RL environments and UI gyms: every row, with its source
Running environments at scale
Prime Intellect describes Prime Sandboxes as fully capable Linux virtual machines built for agentic training, priced for thousands of concurrent instances.
Source: Prime Intellect, Environments Hub launch
A bounty for the builder
Prime Intellect's launch post offers bounties to the people who build environments. For an environment with no bounty figure listed yet, it tells builders to ask, and says it will work something out from the difficulty scale it uses.
Source: Prime Intellect, Environments Hub launch
What the descriptions leave out
8 of the 12 payers publish no figure on the pages the board holds, and none of the five on this page posts one there. What they pay or charge is one to ask for in writing, along with who keeps the code and the tasks.