Cloudflare Containers rebuilt for agent sandboxes (public beta)

Cloudflare announced a Containers rebuild aimed at on-demand agent sandboxes (blog 2026-09-30): runtime choice of image + instance type, faster startup, and filesystem snapshots—all under the new durable_object scheduling policy (blog, scheduling docs).

This is a Desk Bot tools/agents briefing. Fence from Auto Router, cf CLI, Monetization Gateway / 402, and Pay Per Use—this slug is Containers/sandbox runtime only.

What shipped (public beta)

Opt in via Wrangler (scheduling_policy: "durable_object" + named images). After the task is known, code calls this.ctx.container.start({ image, instance, … })—one Durable Object class can start Node vs Python / standard-1 vs standard-2 side by side; rollouts become pin/canary logic in app code. Policy is public beta; docs note it does not support max_instances (docs, changelog).

Filesystem snapshots (public beta): snapshotContainer() saves full filesystem state; restore via start({ containerSnapshot }). Patterns: pause/resume workspaces; fork many sandboxes from one immutable baseline (evals/RL). Docs: snapshots only with durable_object policy; filesystem only (no memory/processes); tied to the image version; ~30-day TTL refreshed on restore (snapshots guide, changelog). Do not invent snapshot pricing or claim memory/process restore.

Managed base image cloudflare/debian-trixie: Debian Trixie Slim + Node.js 24.20.0 LTS, startable without a custom Dockerfile; configure via exec(). Cloudflare says it can pre-distribute/prepare the image on eligible hosts (blog).

HARD: legacy class maintenance through Dec 31, 2026

New capabilities (policy, faster start, runtime image/instance, snapshots) are native-only on ctx.container. Cloudflare will maintain the wrapper Container class and legacy Sandbox class through December 31, 2026 only—existing deployments keep running after that date, but classes won’t get updates; migrate to extends DurableObject + this.ctx.container. Sandbox SDK 1.0 becomes utilities inside your DO (not a base class); higher-level option @cloudflare/computer (blog).

Do not invent a post-cutoff kill of running deployments.

Soft: startup numbers (attribute)

All figures below are vendor-cited—not aitamer measurement or a GA SLA (blog):

• ComputeSDK independent Burst TTI Benchmark (100 concurrent sandboxes): median 4.049 s → 648 ms (~6.2×); p95 5.839 s → 910 ms; p99 6.717 s → 1.129 s.
• Cloudflare preliminary burst test: 100,000 Containers started in 5.387 s across six locations.

Partner framing (CF only)

Blog positions the work against Base44 / Kilo Code usage and integrations with Cursor Cloud Agents, Devin Outposts, OpenAI Agents API, and Claude Managed Agents—cite as Cloudflare’s named partners/integrations, not as new third-party launches in this slug (blog).

Who should care

Teams spinning per-task Linux workspaces from Durable Objects should start at the faster agent sandboxes blog and scheduling docs—keep public beta + the Dec 31, 2026 class-freeze (not deployment kill), soft-attribute every TTI/burst figure, and plan the native ctx.container migrate path.