DeepSeek Paper Details DSec System for Large-Scale AI Agent Training
The infrastructure combines four execution environments with reinforcement-learning workloads to preserve agent state while reallocating computing resources.
DeepSeek has published a paper describing an infrastructure system for training and evaluating AI agents at scale, combining several types of isolated execution environments with resource management designed for long-running interactions.
The paper, titled DeepSeek Elastic Compute (DSec), presents a unified software development kit for four execution backends: function calls, containers, microvirtual machines and full virtual machines. DSec manages the placement and lifecycle of those environments across a cluster.
The sandboxes are designed for agent tasks that require persistent state. They can inspect software libraries, execute commands, call tools and interact with services built for specific tasks. The system also uses memory sharing, resource reclamation and CPU scheduling to increase execution density, while images are loaded as needed from the distributed 3FS file system.
The system was co-designed with a reinforcement-learning framework to decouple stateful rollout execution from preemptible GPU training. That arrangement allows unused resources to be reclaimed while preserving the state of ongoing agent interactions.
The design addresses a workload in which isolated environments may need to be created in large numbers, support different capabilities and remain available across extended sessions. DSec combines those requirements through multiple execution layers rather than relying on a single sandbox type.
Liang Wenfeng, DeepSeek’s founder, is listed among the paper’s authors. The paper lists more than 130 authors.
