Cloudflare Containers revamp: 6x faster, snapshots for AI agents

Cloudflare Containers has existed for a while, but the company previously assumed the containers were configured and deployed by human software developers — assembled, then used to run an application. The spread of AI agents changed that assumption: many agents began using container-based virtualization as their own development, test and execution environment, creating demand for large numbers of container-based virtual machines at once.

The revamp responds to that shift on several fronts at the same time. A new scheduling method drives the start-up improvement; container image and instance size can now be specified in code at run time; new base images remove the need to write a Dockerfile before an agent sandbox can run; and snapshots let a filesystem state be saved and restored. In Cloudflare's own framing, each sandbox's image and instance type can be chosen at run time and everything is controlled from a Durable Object.

The open question is how much of this holds outside the numbers Cloudflare cites. Snapshots are still only in public beta, so reliability and cost of restoring saved environments are not yet settled, and the fast start-up is tied to a new scheduling approach rather than to the underlying hardware. For teams already running agent sandboxes, the appeal is likely to hinge less on raw speed than on whether prebuilt images and snapshots actually remove the setup work agents now generate.