Cua explained: computers, drivers and benchmarks for agents
Deploy Cua: local sandbox, host driver or cloud Fleet
Choose isolation and lifecycle before exposing any computer-use tool.
What you will learn
- Compare the three operating paths
- Plan the image and cleanup
- Keep the network narrow
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
- Fleet and local mode differ in image and credential handling.
- A claim and its pool have different lifetimes.
- Guest isolation does not erase external connections.
Compare the three operating paths
Local Sandbox runs a guest on your hardware. Fleet uses a hosted pool and a claim to reserve a guest. Driver can also target an existing desktop directly. The Sandbox SDK concept is shared, but images, credentials and supported operations differ.
A Linux container guest shares the host kernel, while a full VM has its own guest kernel. Pick according to startup time, OS fidelity and the sensitivity of the task. Neither option eliminates all connections to outside services.
Plan the image and cleanup
An Image describes a starting state; the running sandbox accumulates changes. Fleet uses a published boot artifact and may reject local image-builder layers. Document how dependencies enter the guest before relying on a reproducible launch.
A persistent sandbox survives connection exits, whereas an ephemeral one is destroyed. Fleet pools and claims have separate lifetimes. Delete paid capacity deliberately after testing and export results before destroying a guest.
Keep the network narrow
Driver actions, shell commands and screenshots may reveal sensitive state. Restrict credentials, host permissions and exposed services to the task. A production rollout needs logs, retention rules and an incident owner.
This review did not deploy any path. Run a small restart and cleanup drill on the selected version before making availability or isolation claims.
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 host, local guest or Fleet for the task.
- 2
Document image, credentials and resource lifetime.
- 3
Test cleanup and result export with synthetic data.
Copy-ready example
deployment:
target: local-sandbox-or-fleet
image_revision: record
credentials: scoped
result_export: before-deletion
pool_cleanup: verifyFrequently asked questions
Does disconnecting delete the sandbox?
No. The sandbox docs distinguish disconnect, persistent state and deletion.
Can Fleet run local image-builder commands?
The documented Fleet path expects a prebuilt artifact rather than those local layers.
Sources
- Cua / README.mdSource checked 2026-09-29
- Cua / docs/content/docs/concepts/how-sandboxes-work.mdxSource checked 2026-09-29
- Cua / libs/cua-driver/python/src/cua_driver/fleet.pySource checked 2026-09-29