Reinforcement learning environments
Agent benchmarks that show how environments are built
Last updated /Reinforcement learning environments
A benchmark is useful as a template when it publishes the state, the tasks, and the grader. A leaderboard without those three is a number you cannot buy against. This page reads four public benchmarks that way: tau-bench, OSWorld, ITBench, and SWE-bench. It does not copy their model scores. Where two writeups of the same benchmark disagree on a rate, the rate is omitted on purpose.
The field guide is RL training environments: a field guide. The pillar is reinforcement learning environments: what they are and what real work adds. The live sibling on interfaces is Gymnasium-style APIs for agent environments. The hub page is open RL environment hubs. Checking a file before you trust it, which is a different review from reading a paper, is how to diligence a business dataset before licensing.

Why benchmarks are useful templates
You are not trying to win their leaderboard. You are trying to see how someone else named the world, the task, and the check. If the paper or the repo shows those three, you can ask a vendor for the same three in your domain. If it shows only a percent, you learned a headline. The cards on this page name domain, state, and grader type. They do not show a score.

tau-bench: tools, policies and simulated users
tau-bench is a benchmark of conversations between an agent and a simulated user. The agent has domain tools and policy guidelines. The paper compares the database at the end of the conversation with an annotated goal state, so the wording can vary and the record can still be right or wrong. It also says the simulated user does not see the agent's tool calls. The paper is https://arxiv.org/abs/2406.12045 . Use that as a template for support-style tasks: hidden user goal, written policy, tools, and an end-state check. Do not treat it as a score you can quote for your desk.
OSWorld: real computers and execution checks
OSWorld is a computer environment for tasks on real operating systems and applications. Its paper grades by execution: example-specific scripts read the final state of the machine, including files and the accessibility tree, rather than scoring a screenshot alone. The paper is https://arxiv.org/abs/2404.07972 . The lesson for a buyer is the check, not a pasted success rate. If your task is a desktop or a browser, ask the vendor which file or setting the checker opens.
ITBench: IT operations scenarios
ITBench is a benchmark framework for IT automation tasks. The arXiv paper's abstract says the initial release targets site reliability, compliance and security operations, and financial operations, and that the setup includes scenarios plus evaluation rather than a static quiz. The paper is https://arxiv.org/abs/2502.05352 . Later versions of the writeup do not agree with that abstract on scenario counts and resolution rates, so this page does not repeat those figures. The useful part for an environment is the shape: an IT scenario, the operational signals the agent can see, and an evaluator. Your ticket flow is not their scenario list.
SWE-bench-style coding tasks
SWE-bench styles a task as a real software issue plus a codebase. The paper evaluates by applying the agent's edit and running tests, so the grade is whether the tests that should fail now pass, not whether the patch looked plausible. The paper is https://arxiv.org/abs/2310.06770 . That is a template for work that already has an executable check. It is a poor template for a judgment call with no test. Do not import a coding leaderboard into a support or finance purchase.
Lessons for business-workflow environments
- Name the state the agent sees and the state the grader reads. tau-bench and OSWorld both publish that split in the papers above.
- Publish the task, not only the score. ITBench's paper is cited for its scenario areas, not for a rate.
- Prefer an executable check when one exists, as SWE-bench does with tests. Do not invent a test the job does not have.
- Do not treat any of these percents you may see elsewhere as a promise about your workflow.
We did not run these benchmarks for this page. A card with four names is not a result. If you need records from a company rather than a public paper, that is a different request, and a paper is not a dataset.
Questions people ask
Why are there no success rates on this page?
The papers report rates, and some later writeups disagree. Copying one number would pretend it is stable. Read the paper if you need its table.
Is tau-bench a product we sell?
No. It is a published benchmark. We cite how it grades the database.
Does OSWorld mean we should grade screenshots?
The paper cited above grades the final machine state. A screenshot alone is the weak version of that idea.
Which ITBench number is official?
We do not pick one. The abstract and a later proceedings page do not match. The citation above is the paper. The domains are the part that stayed stable enough to repeat.
Can SWE-bench grade a CRM task?
Only if the CRM task has an executable check of the same kind. A test suite for a code edit is not a pipeline outcome.
Where is the hub page?
It is live and linked above, next to the field guide and the API page.
Describe the data you need
Tell us the use, the records, and the limits. A request is not a promise that a matching dataset is available.
Tell us what data you need