Cua explained: computers, drivers and benchmarks for agents
Cua quickstart: verify one desktop action safely
Use the documented Calculator exercise as a bounded first result.
What you will learn
- Select the target before installation
- Observe, act, verify
- Clean up the test
Before you start
- A disposable target and permitted task
- Basic understanding of host versus guest state
Capture observation, action, policy and final state for one bounded desktop task.
Key takeaways
- Target selection precedes tool selection.
- Verify the application state, not just the agent answer.
- Fleet claim release may not remove pool capacity.
Select the target before installation
Decide whether the action belongs on a personal desktop or inside a disposable guest. Driver operates the desktop where it is installed and targeted. For untrusted tasks, prefer a sandbox rather than testing on a workstation with private data.
The README links to OS-specific Driver installers and a first-app tutorial. Review the install script, platform permissions and supported integration before executing any remote shell command. This article does not run the installer.
Observe, act, verify
The first-app path asks the agent to open Calculator, compute 6 × 7 and confirm the display shows 42. Save a screenshot or accessibility observation before acting, record the action, then read the displayed result.
A natural-language answer of “42” is insufficient: it could have been produced without the application. Verification should refer to the target app state and distinguish a failed click from a correct arithmetic answer.
Clean up the test
Close the target session and remove any disposable sandbox. If using Fleet, follow the provider’s pool cleanup instructions because releasing a claim may leave paid warm capacity.
Keep the trial record short: OS, Cua component version, permissions granted, action, observed result and cleanup. The documented path is a recipe for readers, not an executed test in this review.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Choose a disposable target and inspect the installer.
- 2
Observe Calculator, enter 6 × 7 and verify its display.
- 3
Close the session and confirm resource cleanup.
Copy-ready example
target: disposable desktop
action: enter 6 * 7 in Calculator
verify: displayed result equals 42
cleanup: release claim and delete unused pool if applicableFrequently asked questions
Can I install Driver on my daily desktop?
The project supports host desktops, but a disposable guest is safer for an initial agent trial.
Did EasyAI run the Calculator demo?
No. This is a source-based walkthrough of the upstream tutorial.
Sources
- Cua / README.mdSource checked 2026-09-29
- Cua / libs/cua-driver/README.mdSource checked 2026-09-29
- Cua / docs/content/docs/concepts/how-sandboxes-work.mdxSource checked 2026-09-29