Running AI agents usually forces a bad trade-off. You either isolate them in a sandbox cut off from project data, or you pay to keep full containers idle while waiting on prompts.
Task containers offer a different approach:
- Single-purpose execution: Define a task in your config.yaml to run exactly one command. Once the process exits, the container is removed.
- Full project access: Run tasks in preview environments with access to cloned production data and services instead of working in a blind sandbox.
- Zero idle costs: Billing only covers the exact seconds the command runs.
It gives your agents real infrastructure to complete work, test changes, and spin down without wasting resources.
Check out the full article to see how task containers work under the hood and how to set them up:
Task containers: infrastructure built for AI agents | Upsun
Upsun's task containers run once and disappear, giving AI agents access to production-like data without paying for idle infrastructure
upsun.com
Top comments (0)