DeepSeek researchers report sandbox system running more than 380,000 agent environments at once

DeepSeek researchers report that their production system supports more than 380,000 concurrent sandboxes, the isolated computing environments used to train and test AI agents. In a preprint, they say one deployment spanning around 160 nodes serves about 3 million sandboxes per day and sustains more than 5,000 new sandboxes per second. The system offers several kinds of isolated environments through one software interface. It places them across a cluster, loads data as needed and preserves an agent's work while idle resources are reclaimed. Those capacity figures are the researchers' reported production results.