Compute & Sandboxes
Agent Computer
Persistent cloud computers for agent code execution and remote workflows
Category
Compute & Sandboxes
Isolated computers, execution environments, and managed runtimes where agents can run code or operate software safely.
Editorial scope version 1.0 · effective 3 September 2026
Category listing
2 Compute & Sandboxes Tools
Definition
What belongs in Compute & Sandboxes?
This category covers compute that materially improves an agent’s ability to execute, inspect, resume, or contain work across code, command-line, desktop, and remote-machine tasks.
Includes isolated code, command-line, desktop, and remote-machine environments with agent-oriented provisioning, control, persistence, containment, or observability.
Scope
Inclusion and exclusion criteria
Include when
- Agents can provision, control, or resume an isolated execution environment programmatically.
- The product provides meaningful containment, persistence, observability, or computer-use support.
- Agent execution is a documented first-class workflow rather than an incidental use of generic hosting.
Exclude when
- It is ordinary cloud hosting or container infrastructure with no substantive agent workflow.
- It only wraps a shell command without isolation, state, or control benefits.
Selection guide
How to compare these tools
- What can the agent execute or control?
- What isolation and persistence boundaries are provided?
- Can a person inspect, stop, or resume the environment?
Typical use cases
- Run untrusted or generated code in an isolated environment
- Preserve and resume an agent's execution state across tasks
- Inspect, stop, or recover agent-operated computers
Category evidence
Sources for this category guidance
- Cloudflare Sandbox SDK overview
The Sandbox SDK documentation describes isolated container environments for running untrusted code, with APIs for commands, files, background processes, and exposed services.
accessed 2026-09-03
Comparable facts
Compute & Sandboxes comparison
Unknown values are shown explicitly rather than inferred from marketing copy. Open a profile to inspect its claim-level sources.
| Tool | Classification | Pricing | Interfaces | Deployment | Verification |
|---|---|---|---|---|---|
| Agent Computer | Agent-enabling | Paid | command-line interface, TypeScript SDK, REST API, SSH, browser and VNC | hosted bare-metal infrastructure, private infrastructure (enterprise) | documentation reviewed |
| Cua | Agent-enabling | Freemium | Python SDK, TypeScript SDK, MCP server, CLI | local, Cua Cloud, on-premises | documentation reviewed |