# UseDesktop

> Verifiable RL task environments and trajectory data for computer-use agents.

UseDesktop is a research lab and software platform that turns real desktop workflows into verifiable RL task environments and high-quality trajectory data for computer-use agents. The company supplies frontier AI teams with workflow data packages for post-training, evaluation, and reinforcement learning.

In 2026, UseDesktop is positioned as an RL environment startup for computer-use agents: reinforcement learning environments for agent evals, verifier-backed tasks, reward signals, and high-quality workflow trajectory data.

## What UseDesktop Does

- Captures long-horizon desktop workflow trajectories across browsers, files, spreadsheets, PDFs, portals, and legacy apps.
- Converts workflows into reproducible RL task environments with task goals, UI states, traces, artifacts, outcomes, verifiers, eval splits, failure traces, and reward signals.
- Benchmarks model attempts with pass@k reports, verifier audits, failure taxonomies, and contamination checks.
- Supports two customer paths: teams can use Desktop as end-to-end workflow data infrastructure, or UseDesktop can supply verifiable RL task environments and trajectory datasets with proof.

## Core Pages

- https://usedesktop.com/ - Main landing page for verifiable RL environment and trajectory data infrastructure for computer-use agents.
- https://usedesktop.com/explore - Target workflow classes for verifiable RL task environments, trajectories, verifiers, eval reports, and reward signals.
- https://usedesktop.com/start - Install Desktop and start capturing real workflows.
- https://usedesktop.com/about - Research lab thesis and infrastructure positioning.
- https://usedesktop.com/enterprise - Frontier-team workflow data, RL task environment, and post-training pipeline.

## Key Concepts

- CUA: computer-use agent.
- RL task environment: a reproducible task package with state, actions, outcomes, verifier logic, reward signal, and train/eval split.
- Agent evals: model runs against verified computer-use tasks with pass@k, failure modes, verifier results, and reward signal analysis.
- Trajectory data: human or agent action sequences with screenshots, UI states, artifacts, failures, corrections, and final outcomes.
- Verified data: workflow data with active testing, verifier audits, difficulty calibration, model comparisons, and contamination checks.
- Real workflow data: operational desktop work performed by people solving actual tasks, not synthetic benchmark-shaped tasks.

## Contact

- Website: https://usedesktop.com
- Email: support@usedesktop.com
- Book a call: https://cal.com/seungju-chae