# Docker in sandboxes, larger sandboxes, and longer runs

> Cloud sandboxes run Docker and Compose, agents can move to a larger sandbox as they work, and code workflows can ask for up to 4 hours.

Nov 17, 2026. Topics: Workflows, Agents, Pricing. https://www.generalinput.com/changelog/2026-11-17-docker-in-sandboxes

Every cloud sandbox now has Docker. An agent can start a database in a container, build an image, or bring up a whole stack with Docker Compose, and a code workflow can do the same from its script. The first `docker` command starts the engine, so a sandbox that never uses it carries no extra load.

An agent, in a chat or an agent workflow, can move its own sandbox to a larger machine when the work needs it, such as a heavy stack like Elasticsearch or a big build, and back once it's done, with its workspace files. Small containers and light stacks run on the standard one.

A code workflow can say what its sandbox needs in a `sandbox.json` file: Docker, a larger machine, and more time. That makes jobs like a scheduled security scan of your own app possible as [workflows](/developers).

### How it works

- **Docker and Compose.** Docker Engine with the Compose and Buildx plugins. A workflow that needs them says `"requires": ["docker"]`, so it only runs where Docker does. A self-hosted server's local sandboxes can't.
- **Large sandboxes.** 8 vCPUs and at least 8 GiB of memory, for a code workflow with `"size": "large"` or an agent that switches itself. A conversation left idle starts standard again.
- **Large time costs 160 credits an hour.** That is $1 an hour, by the second, twice the standard rate of 80 credits. Each size has one rate, whichever machine runs it. See [pricing](/pricing).
- **Longer runs.** `timeLimitMinutes` goes up to 240. Without it a run stops at 60 minutes, as before.
- **Never without what it asked for.** Before a run's code starts, its sandbox is checked for what it requires, and a run that can't get that there moves to one that can.