DeepSeek DSec: Sandbox Infrastructure for Agentic AI Training
Original: DeepSeek Elastic Compute (DSec)
Why This Matters
Shows the infrastructure scale required to train capable agentic AI systems at production level.
DeepSeek published a paper on DSec (DeepSeek Elastic Compute), a production-scale sandbox platform supporting ~3 million sandboxes per day across ~160 nodes, handling 380,000+ concurrent environments for large-scale agentic LLM training and reinforcement learning workloads.
DeepSeek's arxiv paper (arXiv:2609.22978) details DSec, a sandbox infrastructure built to support large-scale agentic training with LLMs. The core problem: agentic workloads spin up sandboxes in massive bursts, require stateful execution across long interactions, need heterogeneous isolation levels, and pull from vast image libraries with minimal reuse — a combination that breaks conventional single-runtime sandbox designs.
DSec addresses this by exposing four backend types — FnCall, container, microVM, and full-VM — through a single unified SDK. The platform handles placement, lifecycle management, and layered environment composition using independently versioned image layers. For high-density execution, it combines memory sharing, memory reclamation, and CPU scheduling. Image data is loaded on demand via 3FS (Fire-Flyer File System), DeepSeek's in-house distributed filesystem.
Critically, DSec is co-designed with DeepSeek's RL training framework. It decouples stateful rollout execution from preemptible GPU training, coordinating sandbox lifecycles with training runs to preserve rollout state while reclaiming idle resources. It also includes mechanisms to detect and mitigate agent misbehavior such as reward hacking. A single production unit spans roughly 160 nodes, serving about 3 million sandboxes per day with over 380,000 concurrent environments.