DeepSeek DSec Sandbox Infrastructure Optimizes Scalable Agentic Training

DeepSeek has introduced DeepSeek Elastic Compute, commonly referred to as DSec, a specialized sandbox infrastructure engineered to address the complex demands of large-scale agentic training for large language models. Published as a preprint on arXiv and under review for the ACM SIGOPS ATC 2026 conference, the system represents a significant advancement in distributed execution environments tailored for autonomous AI workloads. Agentic training requires isolated, stateful execution environments where models can inspect code repositories, invoke external tools, run system commands, and interact with specialized services. These operations generate unpredictable bursts of sandbox instances that demand heterogeneous isolation levels, long-term state retention, and efficient handling of large image libraries with minimal reuse. Recognizing that a single runtime cannot meet these diverse requirements, DeepSeek designed DSec as an elastic execution platform. The architecture provides a unified software development kit that abstracts multiple backend environments, including function calls, containers, microvirtual machines, and full virtual machines. At its core, DSec manages cluster-wide placement and lifecycle coordination, assembling environments from independently versioned layers. The system optimizes resource utilization through advanced memory sharing, aggressive reclamation protocols, and dynamic CPU scheduling to support high-density workloads. Additionally, it leverages the Fire-Flyer File System, or 3FS, a cluster-wide distributed storage solution that enables on-demand loading of sandbox image data, significantly reducing network and storage bottlenecks. The platform was explicitly co-designed alongside DeepSeek’s reinforcement learning framework to ensure seamless integration between inference and training pipelines. By decoupling stateful rollout execution from preemptible GPU training cycles, DSec preserves critical rollout states while dynamically reclaiming idle compute resources. The architecture also incorporates safeguards to mitigate common agentic failures, such as reward hacking and environment misuse. Production benchmarks demonstrate substantial scale and efficiency. A single DSec deployment unit, comprising approximately 160 compute nodes, handles over three million sandbox instances daily. The system currently supports more than 380,000 concurrent sandboxes and sustains a provisioning rate exceeding 5,000 instances per second. Evaluation metrics indicate that DSec successfully reduces environment initialization and image distribution latency while maintaining stable, latency-sensitive performance even under heavy resource overcommitment conditions. The architecture reflects a broader industry shift toward specialized infrastructure that bridges the gap between static cloud sandboxes and the dynamic, stateful requirements of autonomous AI agents. By streamlining execution environments and tightly coupling them with reinforcement learning workflows, DeepSeek aims to accelerate the development and evaluation of capable agentic systems at an operational scale previously constrained by infrastructure limitations.