Docker Cloud Sandboxes Let Developers Seamlessly Scale AI Agent Tasks from Laptop to Cloud

Docker has launched Cloud Sandboxes, a cloud-hosted execution environment built on Docker-managed infrastructure that mirrors its earlier local sandbox offering. The service enables developers to run AI coding agents in secure, hardware-isolated microVMs and move workloads between a developer's laptop and cloud infrastructure without rewriting deployment logic.

Moving from Local Bursts to Cloud Duration#

The shift reflects a practical reality as AI agents tackle increasingly demanding tasks. When agent sessions lasted minutes, the chief concern was whether the model could maintain context. Now that agents run for hours at a stretch, the limiting factor is the hardware itself: personal laptops sleep, throttle on battery, and lose connectivity when moved.

Cloud Sandboxes let developers initiate work locally and promote it to the cloud as needed—or hand off tasks to run unattended while they disconnect. The same CLI interface and isolation model apply to both environments, making the transition transparent.

Single-Command Portability#

Moving a sandbox between environments requires just one command:

sbx move my-project --to cloud

This captures the sandbox's state—including its filesystem—and replicates it on the other side, preserving all progress and configuration. Developers can now orchestrate dozens of parallel agent tasks, each running in its own isolated instance for extended periods.

Kits as Standard OCI Images#

Docker is also advancing its sandbox ecosystem with Kits v3—pre-configured sandbox templates now packaged as standard OCI images. This change means Kits work like any Docker image: they can be pulled, built upon, and combined, lowering barriers to composing complex agent environments.

Questions About Threat Surface#

Early feedback highlights trade-offs. While sandboxing isolates agent execution, real-world agents must reach external services—package repositories, cloud APIs, specialized tools. Security researchers note this creates a new attack surface: an agent confined to a sandbox could still compromise services it's permitted to contact. Some advocates argue that capability-based access controls, rather than traditional process isolation, may better constrain what agents can request and from whom.