Which Docker Host Passes Your Agent Recovery Drill?
Use Docker Cloud Sandboxes for jobs that fit their configured lifecycle. Compare replacement-container recovery on a self-managed VPS or managed host when sessions and artifacts must survive.
Measure each kind of restored state independently
For a Claude Agent SDK deployment, plant three markers in a disposable workload: a unique fact in the transcript, a different fact in project memory, and a file with a recorded checksum. Force recovery by starting a fresh test container from the same image with the candidate host's intended persistence configuration. Resume the session, ask for each fact separately, then verify the file checksum. The SDK's SessionStore adapter can persist transcripts across hosts. Memory files and working artifacts need storage of their own. Start timing when replacement begins and stop when the last required check passes. In an illustrative drill, a transcript recovered in 20 seconds and a missing artifact at the five-minute cutoff amount to incomplete recovery. Compare hosts using the same markers and cutoff.
Hosting the Agent SDK - Claude Code DocsBudget one replay against the sandbox lifetime
Docker documents a one-hour default expiration for cloud sandboxes. At expiration, the service stops sandboxes that support resumption and deletes the others. For a bounded workload, design for one replay by measuring how long it takes to restore the last durable checkpoint, redo discarded work, and finish exporting the result. An illustrative 42-minute job with 8 minutes of lost work and 5 minutes of restore time leaves only 5 minutes under a 60-minute lifetime when the whole sequence shares one clock. Enter measured end-to-end job time including final export and use the configured timeout for your trial. For a VPS or managed host, compare the replacement drill separately and price the storage and operational work needed to meet the same recovery target.
Use cloud sandboxes | Docker DocsStart with the expiration policy
Docker Cloud Sandboxes expire after one hour by default. At expiration, the service stops sandboxes that can be resumed and deletes the others. Check which outcome applies to your configuration before handing it a job. Start with the longest expected run, include time waiting for human approval, and decide where results will be available after expiration.
For an illustrative test, give a fresh sandbox a 15-minute document conversion, a 10-minute approval window, and an external destination for the output. Run it twice and have the reviewer retrieve both results after expiration. A missing output or a job still waiting for approval at cutoff fails the test. If the workload regularly needs that longer window, test a different expiration setting or move it to durable hosting.
Use cloud sandboxes | Docker DocsTest the workspace on each candidate
Docker local sandboxes can mount host paths. Docker Cloud Sandboxes have no access to those paths, so plan to transfer files or clone the repository inside the sandbox. That distinction should decide whether a laptop pilot is representative. A job that reads a local fixture or private repository can work on a laptop and fail immediately on a fresh cloud run.
Run the same sample on Docker Cloud and any E2B or Daytona offering you shortlist. Use a private test repository, a fixture absent from the repository, and a scoped test credential. Record how each environment receives the inputs, where it writes the result, and what a new run can recover. Verify each vendor's expiration, storage, and credential handling separately. Accept a candidate only if a second operator can reproduce the run from the documented inputs.
Compare local and cloud sandboxes | Docker DocsCompare the cost of passing the drill
Hetzner lists Regular Performance cloud servers starting at €5.49 per month excluding VAT. Treat that as a starting compute rate. Price the region and machine that passed your trial, then add storage, backups, model charges, and operator time. Get a managed quote for the same runtime mix and number of workspaces, and check whether it bills per instance, seat, or usage unit.
In an illustrative comparison, assume €12 per month for the tested VPS, €4 for backup storage, and 30 minutes of monthly upkeep at €60 per hour. The total is €46 before tax and model usage. Against a hypothetical €35 managed quote with equivalent backup and model billing, the managed option leads by €11. If your existing platform absorbs upkeep and the same recovery test passes, the VPS subtotal falls to €16.
Cloud-hosting provider for developers & teams