Get Started

RL Environments: Enterprise Atlas

Training agents to complete long-horizon work across real enterprise systems and desktop environments.

Request Samples
A dark office seen through glass at night, with colleagues meeting at a table and one person working at a desk

Enterprise Atlas: Frontier data for frontier models

Data from experts and real companies

We create environments with both real enterprise and expert-generated data, retaining enterprise fidelity, complexity, and diversity.

Model stumping, verified results

Expert-designed long-horizon, cross-functional tasks and rubrics, with verifiable post-training results.

Real applications, contained environments

MCP servers, CLI, and GUIs that behave like real-world systems in a stable environment.

Tool use environments

Real World

Environments built with a real company's operating history, anonymized and rehydrated to retain coherence and complexity, safe for training.

Business (Mobile game studio):

  • 76+ scenarios
  • 10 systems
  • 500+ tool calls on super long-horizon tasks

Example business functions

  • Product management
  • Engineering ops
  • Marketing & GTM
  • Finance

Real operating history

  • Employees: 51
  • Time span: 2023–2026
Two colleagues looking at a computer screen together in an office

Advanced

High-fidelity reconstruction of real enterprise environments, requiring agents to solve non-contrived, end-to-end tasks.

Businesses:

  • Accounting firm
  • mortgage broker
  • logistics company

Example Functions: Finance, accounting, loan operations, compliance, customer support

Results (Qwen3.8-27B, OPD + GSPO):

  • +13pp Toolathlon
  • +3pp AutomationBench
Two engineers reviewing code on a monitor

Train agents to solve end to end enterprise tasks using real tools

2000+

Tasks

4

Enterprise Environments

870K

Artifacts

Computer use environments

Built for OSWorld 2.0 + Agent Last Exam

Professional Office

Long-horizon computer use tasks grounded in real professional artifacts across 11 domains.

  • 1,000 unique tasks
  • 100+ avg steps per task
  • Verified by a rubric-based agent judge
  • Compatible with Linux, Windows, and MacOS


Domains: Engineering, STEM Research, Design & Creative, Multimedia, Finance & Accounting and 6 more.

Close-up of hands typing on a laptop with code on the screen

Built for OSWorld 1.0

OSWorld Advanced

OSWorld-style tasks weighted toward the hardest difficulty tiers.

  • 370 unique tasks
  • 70% hard or very hard difficulty
  • 40% multi-app workflows (vs. 27% in OSWorld)
  • Compatible with Linux, Windows, and MacOS

Categories: Multi-app, Chrome, Office, VS Code, OS, and Mail

Two engineers reviewing code on a monitor

Try Scale RL Environments

Book A Demo
Motion-blurred commuters walking across a tiled station floor, seen from above