Provision an AI worker once, then dispatch tasks to it whenever you need — pay only per use. Each subagent runs in its own persistent cloud sandbox with its own memory.