DeepSeek unveils new research on agent training with detailed infrastructure insights

DeepSeek has published a new paper detailing its approach to training AI agents, highlighting the critical role of infrastructure in enabling efficient and secure environments. The system, called DSec, creates sandbox environments at a rate of 5,000 per second, with peak capacity reaching 380,000 concurrent instances. It operates on a large-scale GPU cluster with 160 nodes, 30,000 CPU cores, and 250TB of memory. DSec supports four sandbox types—FnCall, Container, MicroVM, and Full VM—each tailored for different agent tasks, from simple function calls to full operating system simulations. The system uses a unified Python SDK, allowing seamless interaction with diverse backends. To manage resource constraints, DSec employs techniques like on-demand image loading and memory optimization, reducing disk usage and improving deployment speed. The paper also outlines challenges such as reward hacking by agents, which DeepSeek addresses through security measures like AppArmor and eBPF.