Cloudflare has redesigned its Containers service around the short-lived, task-specific computers used by AI agents. Application code can now select a container image and instance size at runtime, while filesystem snapshots let a workspace pause and resume with its files intact.
An independent ComputeSDK benchmark measured median startup at 648 milliseconds, down from just over four seconds. Cloudflare describes that as more than a sixfold improvement and says preliminary burst tests created hundreds of thousands of containers in seconds. Those figures reflect benchmark and early test conditions rather than guaranteed startup for every workload.
Each container retains a Durable Object, a persistent controller that gives the environment an identity and manages its lifecycle and outbound traffic. A new scheduling policy moves image and resource choices from deployment configuration into the application handling the task. That allows one agent to request a small workspace and another to request a GPU-capable environment without maintaining a separate preconfigured application for every combination.
The design targets coding agents, evaluations and reinforcement-learning systems that need isolated Linux environments, package managers and test tools. Snapshot support is in public beta, and the updated control model is being carried into Sandbox SDK 1.0. Teams still need to size resources and manage the security implications of code-running agents.