DSec: How DeepSeek builds 3 million agent environments a day

Training an agent is not about feeding data to a network. The agent must write code, compile it, open a browser, and sometimes install an operating system inside a live environment. Every step alters the state of that environment, which can break at any moment. Each training round therefore requires a fresh, clean sandbox that is discarded after use. The Chinese tech outlet IT之家 reports on DeepSeek's system, DSec (DeepSeek Elastic Compute), signed by co-founder Liang Wenfeng.

Scale is the core challenge. DSec generates over 5,000 sandboxes per second, totaling 3 million a day. Peak concurrent load reaches 380,000 environments. A single cluster supports this, comprising about 160 nodes, 30,000 CPU cores, and 250 TB of RAM. Through resource oversubscription and dense packing, one node holds 3,200 containers or 800 micro virtual machines simultaneously.

The problem: a system for each of 3 million sandboxes

The bottleneck is not scheduling, but environment construction. The classic Docker approach packs a base image, a working directory, and a toolset into one complete image. This works at small scale. In DSec, the container backend used 11,266 base images and 102,171 working directories in total. 67.8 percent of sandboxes require at least one extra layer, containing a working directory or tool package, on top of the base image. With such variety, updating a single package would force a rebuild of every image that contains it.

The solution splits the environment into three independently versioned, read-only layers in the EROFS format: base image, working directory, and tool package. Overlayfs joins them at sandbox startup. A package update then affects only its own layer. The second component is Chronus, which mediates communication from inside the sandbox to the training framework. This allows the framework to track the agent's stage and determine the appropriate feedback signal.

Why this matters for every agent model

The V4.1-Flash model card confirms that this was the focus of the training run. The post-training recipe remained standard. All significant changes occurred in the data pipeline, specifically in the mass automatic synthesis of tasks and agent environments. Agent quality now depends on the background environment factory, not on another trick in the learning algorithm. DSec is DeepSeek's answer to producing 3 million such environments daily without building a data center that chokes on itself.