Of the companies on our list, Appen is the only one whose own pages describe a ready-made coding set: 500 coding tasks, every one with its own automatic verifier. Others, including Surge AI, Toloka, Bespoke Labs and Scale AI, build environments, tasks or verifiers to order for AI labs. Coding is also where open environments are most plentiful, so many teams start on a free hub and pay for what it lacks.
Which companies build RL environments for coding agents?
The table sticks to what each company says on the page we read. Where a company does not single out coding, the table says so rather than guessing.
| Company | What its pages describe | Coding-specific detail published |
|---|---|---|
| Appen | Ready-made RL tasks with automatic checks | Yes: a set of 500 coding tasks, each with its own verifier |
| Surge AI | RL environments, rubrics and verifiers for AI labs | Not singled out |
| Toloka | Auto-verifiable tasks for evaluation and RL | Not singled out |
| Bespoke Labs | RL environments plus the engine, sandboxing and execution to run them | Not singled out |
| Scale AI | Enterprise tool-use and computer-use environments | Not singled out (its published sets are enterprise and computer use) |
| Prime Intellect | The open Environments Hub and prime-rl training code | Shared environments of many kinds |
Companies named in this table may pay us a fee if an introduction we make leads to work with them. That does not decide who is listed or the order.
“Not singled out” does not mean a company cannot do coding work. It means its own page, as we read it, does not describe a coding set, so we do not claim one. Ask directly.
What does a coding RL environment contain?
The parts are the same as any environment, but coding gives them a particular shape.
- The sandbox is a container holding a repository at a fixed commit, with its dependencies installed, that resets to the same state for every attempt.
- The task is an instruction such as “fix the failing test” or “add this endpoint”, written against that repository.
- The tools are a shell, a file editor and sometimes a browser for documentation.
- The grader is usually a hidden test suite. If the right tests pass and nothing else breaks, the task is done.
That grader is why coding is popular for RL. Tests are deterministic and cheap to run, so the reward is clean. It is also why coding environments are the easiest to game.
How do coding agents game the grader?
The classic shortcut is editing or deleting the failing test instead of fixing the code. Others include special-casing the exact inputs the tests use, or catching every error so nothing visibly fails. A good coding environment keeps its tests out of the agent’s reach, checks that existing tests still pass, and is tested against deliberately wrong patches before delivery. Our post on how graders catch reward hacking covers this in more depth.
What are the best RL environments for LLM agents that write code?
The best one is the one whose tasks look like your agent’s real work. Three questions sort the options:
- Which languages and frameworks? A set built on one language teaches little about another.
- How realistic are the repositories? Toy repos are cheap to build but do not teach navigation of large codebases.
- Has the task set been public? Public coding tasks may already be in a model’s training data, which inflates scores without teaching anything.
Ready-made set or built to order?
A ready-made set, like the one Appen describes, is quicker to start on. Other buyers may have the same tasks. Built-to-order work is slower and dearer but can target your stack. Epoch AI’s FAQ puts exclusive arrangements at about four to five times the cost of shared ones.
What does a coding task cost?
No company we list publishes a price. Epoch AI’s interviews put one task with its grader at “$200 to $2000 mostly”, with $20,000 rare. Coding tasks with automatic tests tend to sit toward the cheaper end of what drives that range, because the grader is code rather than a rubric or an expert reviewer. Our task costs page multiplies the range out for several task counts.
How should you compare builders?
Ask each for a small sample: tasks, the repository snapshot, the hidden tests and the wrong patches each grader was tested against. Then ask for a priced pilot in your own language and framework. When you compare RL environment companies this way, the differences show up in the graders long before they show up in the price.
For coding agents, the test suite is the teacher, so buy the tests as carefully as the tasks.