CLOUDFLARE has rebuilt its Containers offering to better support on-demand, fast-starting agent sandboxes. The update introduces a durable_object scheduling policy that lets the agent code decide which sandbox image and compute resources to use at runtime, rather than locking those choices in at deploy time. As a result, sandboxes can be created for each task, start more than six times faster, and support filesystem snapshots in public beta, enabling workspaces to be saved and restored across sessions.
The redesign centres on treating the Durable Object as the controller of its attached Container. This allows sandboxes to start from a prepared system image, and to tailor toolchains (for example Node[.]js vs Python) and compute size (standard-1 vs standard-2) per task. Cloudflare cites substantial speed gains: a median startup time drop from just over four seconds to about 648 milliseconds, with bursts reaching hundreds of thousands of containers quickly.
Two new capabilities are highlighted: runtime image and instance selection inside the Durable Object, and native filesystem snapshots, which support continuing work from a saved state and sharing a common baseline for evals or experiments. The piece also notes that this approach eliminates the need for multiple pre-deployed applications to cover different sandboxes, since environment decisions are made by code at request time.
In addition, Cloudflare plans to ship a ready-made image, cloudflare/debian-trixie, to streamline initial starts, and to maintain compatibility with existing APIs while migrating toward the new durable_object model.