Computer Use Environments
Train agents to navigate and operate realistic desktop and web environments, including macOS- and Windows-like systems.
RL Environments
Train and evaluate agents to excel at long-horizon, professional workflows.
Features
Made for Agent Training
Real Applications, Rebuilt for RL
Simulated APIs, MCP servers, and GUIs that behave like the systems agents actually use.
Expert-Curated Artifacts
Files (PDFs, CSVs, PPTs, etc.), data, and state settings that reflect real professional workflows, sourced from real business settings, capturing natural messiness.
Hard Tasks with Verifiable Outcomes
Expert-designed objectives, rubrics, and automated verifiers that produce stable training signals for programmatic process and outcome evaluation.
Easy Integration and Environment Control
Compatible with standard environment interfaces, supporting trajectories, resetting state, examining rewards, and more.
Overview
Inside Scale AI's RL Environments
Built to Advance Agent Capabilities
Scale AI's RL Environments are simulated collections of realistic applications designed to train and evaluate agent behavior.
Systems for Stronger Learning Signals
They mirror real computer and API-based systems, supporting rich logs and application state that can be used for programmatic and rubric-based evaluation.
RL-Ready Expert Data
Each environment combines simulated interfaces with expert-curated data and evaluators to produce reliable learning signals. Scale also provides analysis on complexity (pass@k) and anticipated training gains.
Capabilities
Coverage That Matches Production-grade Systems
MCP / Tool Use Environments
Train agents to reason over and use real tools via MCP servers — Slack, HubSpot, Linear, and more.
Available as
Environment Types
Three types of simulated environments — each designed to match the systems agents operate in production.
Web Apps
Browser-based SaaS and productivity apps.
Desktop VMs
macOS- and Windows-like operating system environments.
MCP Servers
Real enterprise tools — Slack, HubSpot, Linear, and more.
Try Scale RL Environments
