RL environments to train AI agents on real work
Every task has a proven answer. Every reward is earned.Every task has a proven answer. Every reward is earned.
Docker · HTTP API · MCP · OpenEnv · Harbor · Browser GUI
OpsDesk Workforce
The HR and staffing back office of a shift-based employer, rebuilt as an RL environment: rota software, mail, team chat and calendar on one state. Computer-use agents work in the browser; tool-use agents call the same functions over HTTP or MCP.
English edition (German employer and labour law): requests, documents, mail, chat and app texts in English; German law, German names and euros.
Inside every Vencur environment
- The software
- Screens, APIs and tools a person would use.
- The rules
- Laws, house rules, approvals.
- The people
- Sick calls, swap requests, a works council.
- The reward
- A checker that reads the end state, not the agent's report.
How a run is scored
Checked against the app's end state, never against the agent's report.
| The run | Score |
|---|---|
| Solved | Pass |
| Solved another valid way | Pass |
| Almost solved | Fail |
| Reports success, didn't do it | 0 |
| Did nothing | 0 |
Difficulty
On the expert tier, Claude Opus 5.5 passes 2 of 20 attempts, against 88% one level below.
| Tier | Use | Claude Opus 5.5, passed attempts |
|---|---|---|
| StandardTraining | Training | Not part of this measurement |
| HardTraining | Training | 23 of 2496% |
| AdvancedTraining and evaluation | Training and evaluation | 21 of 2488% |
| ExpertEvaluation | Evaluation | 2 of 2010% |
Higher tiers add changes in the middle of a task, information that only arrives when the agent asks, and duties nobody mentions in the request.
Measured on the German edition on 13 November 2026; not yet measured for the English edition. Blind runs through the tool interface, no references and no hints, one request per attempt. Every attempt is scored deterministically against solver-proven references; no LLM judge. A pass means every rule, coverage, message, duty and honesty check is right and the plan costs at most 0.5% more than the proven optimum.
How it works
-
01
Sample
5 tasks. Run your own agent and try to break the checker.
-
02
Pilot
25 fresh tasks for $6,000, half on order and half on acceptance, delivered within 5 working days.
-
03
Licence
Tasks by tier, for evaluation or training.
-
04
Refresh
New, harder tasks every quarter.
Custom environments
Your workflow, replicated: we scope it with you, build it on the same engine and prove it with the same checks.
Built for labs, data vendors and agent teams
-
Frontier labs
Train and evaluate on work that has a right answer: calibrated tiers, solver-proven references and full transcripts in the Proof Report.
-
Data vendors
Partner Licence: license our environments on to your lab customers at a flat fee per task.
-
Agent companies
Test your agent against real rules before it acts in production.
Frequently asked questions
What is an RL environment?
A reinforcement-learning environment is a working replica of a job: the software, the data, the rules and the people an agent deals with, plus a reward that says how well it did. Vencur's rewards come from a checker in code, not from a language model.
What is OpsDesk Workforce?
Our first environment: the HR and staffing office of a shift-based employer under German labour law, in an English edition (German employer and labour law). Every task ships with a solver-proven answer and a checker we try to break before you do. See OpsDesk Workforce
How do you verify a reward?
A deterministic checker reads the end state of the app: valid state, hard rules, coverage, cost against a solver-proven reference, notices and stability. Honesty is checked on the agent's final status report: six fields, each compared exactly with the end state; free text is not judged for honesty. No LLM judge. How we verify rewards
Can agents game the checker?
No shortcut pays. A run that reports success without doing the work scores 0, so does a run that does nothing. An almost-solved run fails; a run solved another valid way passes.
How hard are the tasks?
There are four tiers: standard, hard, advanced and expert. On the expert tier, Claude Opus 5.5 passes 2 of 20 attempts, against 88% one level below. Measured on the German edition; the English edition is not yet measured. See the benchmark
Which formats do you deliver?
Docker (offline, amd64 and arm64), HTTP, MCP, OpenEnv and Harbor. Other harnesses on request. A browser GUI for computer-use agents runs on the same state.
What does it cost?
Free sample: 5 tasks on GitHub. Run your own agent and try to break the checker. Pilot: 25 fresh tasks for $6,000, half on order and half on acceptance, delivered within 5 working days. Licence after the pilot: from $150 per task (training, from 300 tasks) to $400 per task (held-out evaluation for a named lab). Exclusive licences on request. Pricing
Can data vendors resell Vencur environments?
Yes. Under a Partner Licence, data vendors license our environments on to their lab customers at a flat fee per task. Pricing
Who owns the data and the licence?
All code and data are Vencur's own, and every task uses synthetic data only. You get a licence to use them: non-exclusive by default, exclusive on request. You own the models you train.
Can you build our workflow?
Yes. We build custom environments for other operations work, and OpsDesk Workforce tasks on your own rules or anonymised data, all checked the same way. Custom environments
How do we start?
Get the 5-task sample, run your own agent and try to break the checker. Then order a pilot of 25 fresh tasks.
Tell us what your agents need.
Request a sample or a pilot. We'll reply by email.
Thanks. We'll reply by email.