Cloudflare just made its Container platform start almost instantly, specifically so AI agents can spin up disposable Linux sandboxes on demand.
Cloudflare rearchitected Containers around a new durable_object scheduling policy that lets application code pick an image and instance type at runtime instead of locking those choices in at deploy time. The company says this made startups more than 6x faster, and it points to an independent ComputeSDK benchmark showing median cold starts drop from just over four seconds to 648 milliseconds. Cloudflare also shipped filesystem snapshots in public beta, so a sandbox can be paused, saved, and restored later, even days after the original task ended. Each Container still runs behind its own Durable Object, the small stateful controller that manages its lifecycle, but that controller can now swap images on the fly without a separate deployment or rollout process.
The old model meant provisioning a distinct application and namespace for every combination of toolchain and sandbox size, then routing tasks to the right one and waiting on a global control plane to coordinate placement. For coding agents, eval harnesses, and reinforcement-learning systems that spin up sandboxes constantly and briefly, that overhead added up fast. Cloudflare is explicitly chasing agent workloads here, and it name-checks integrations with Cursor, Devin, and the OpenAI Agents API as evidence the market is already moving that way.
The 648-millisecond figure is Cloudflare's own framing of someone else's benchmark, not a number outside users have reproduced at scale, so treat it as a best case rather than a guarantee.