DeepSeek details DSec for agent sandboxes at production scale

A 31-page report posted September 19 describes DeepSeek Elastic Compute, or DSec, for agent training and evaluation across function-call, container, microVM and full-VM sandboxes, managed through one Python SDK with lifecycle, resource and network controls. The authors say one roughly 160-node production unit serves about three million sandboxes a day, while production supports more than 380,000 concurrent sandboxes and over 5,000 creations per second. On-demand image loading and composable layers avoid moving full images, while a reclamation mechanism cut time-integrated memory use 21.2% in evaluation. These are team-reported deployment and benchmark figures from an arXiv preprint, not independent replication. The broader signal is that agent scale now depends on security, scheduling and infrastructure economics as much as model quality.