DeepSeek Unveils DSec: Scalable Sandbox Infrastructure for Agentic AI

Overview

DeepSeek has introduced DeepSeek Elastic Compute (DSec), a production-grade sandbox infrastructure designed to support large-scale agentic AI training and evaluation. The system addresses the complex computational demands of modern large language models (LLMs) that require isolated, stateful execution environments.

Unlike traditional sandbox solutions, DSec operates as an elastic platform capable of handling massive bursts in workload while maintaining high-density performance. A single production-scale unit of the infrastructure spans approximately 160 nodes, serving around 3 million sandboxes daily and sustaining over 5,000 sandbox creations per second.

The technical paper detailing DSec was submitted to arXiv on September 19, 2026, under the title "DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale." It represents a significant advancement in distributed computing for AI agents.

Technical Architecture and Capabilities

DSec provides a unified SDK that exposes multiple sandbox backends, including FnCall, container, microVM, and full-VM environments. This multi-backend approach allows the system to accommodate heterogeneous functionality and isolation requirements inherent in agentic workflows.

The platform coordinates placement and lifecycle management across the cluster, ensuring efficient resource allocation. Key architectural features include:

• Layered Environment Composition: DSec composes environments from independently versioned layers, reducing redundancy and accelerating setup times.
• Resource Optimization: The system combines memory sharing, reclamation, and CPU scheduling to achieve high-density execution without compromising stability.
• On-Demand Data Loading: Image data is loaded on demand from the Fire-Flyer File System (3FS), a cluster-wide distributed filesystem that mitigates storage bottlenecks.

"Supporting agentic workloads requires an elastic execution platform rather than a single sandbox runtime," the authors note in their submission. "These workloads create sandboxes in large bursts, span heterogeneous functionality, retain state across long interactions, and draw from large image corpora with limited reuse."

The integration with 3FS is particularly critical for performance, as it allows the system to distribute the load of image retrieval across the cluster rather than relying on centralized storage solutions.

Integration with Reinforcement Learning Frameworks

A distinguishing feature of DSec is its co-design with DeepSeek’s reinforcement learning (RL) framework. This close integration allows the system to decouple stateful rollout execution from preemptible GPU training, a common challenge in large-scale AI development.

The platform coordinates sandbox lifecycles with training processes to preserve rollout state while reclaiming idle resources. This synchronization ensures that the computational cost of maintaining agent environments is minimized without interrupting ongoing training tasks.

The architecture also includes mechanisms to mitigate agent misbehavior, such as reward hacking. By enforcing strict isolation and monitoring within the sandbox environment, DSec helps maintain the integrity of the training process.

Key benefits of the RL-coordinated design include:

1. Preservation of stateful rollouts during training pauses or resource reallocation.
2. Efficient reclamation of idle sandbox resources to free up capacity for new tasks.
3. Enhanced stability in high-density overcommit scenarios, where multiple agents compete for limited resources.

This tight coupling between inference-time execution and training-time optimization sets DSec apart from general-purpose sandbox solutions that do not account for the specific temporal dynamics of agentic RL workflows.

Production Performance and Evaluation

In production deployments, DSec has demonstrated remarkable scalability and efficiency. The system currently supports over 380,000 concurrent sandboxes, a figure that underscores its ability to handle the massive parallelism required by modern agentic AI systems.

The evaluation reported in the arXiv paper highlights several performance improvements achieved through DSec’s design:

• Reduced Setup Overhead: The use of layered composition and on-demand loading from 3FS significantly decreases the time required to initialize new sandbox environments.
• Improved Memory Efficiency: Memory sharing and reclamation strategies allow the system to pack more workloads onto each node, maximizing hardware utilization.
• Latency Resilience: The platform maintains latency-sensitive performance even under high-density conditions where resources are heavily overcommitted.

The research team, comprising over 130 authors from DeepSeek, notes that their evaluation and deployment experience confirm these theoretical advantages. The system’s ability to sustain such high throughput while preserving stateful interactions is critical for the next generation of agentic AI applications.

The paper, spanning 31 pages and 13 figures, represents a substantial expansion from an earlier two-page extended abstract that underwent first-round review for the Operational Systems Track of ACM SIGOPS ATC 2026. This peer-reviewed pedigree adds credibility to the reported metrics and architectural decisions.

Implications for Agentic AI Development

The introduction of DSec marks a significant step toward democratizing large-scale agentic AI training. By providing a robust, elastic infrastructure that can handle the complexities of stateful, tool-using agents, DeepSeek addresses one of the key bottlenecks in the field.

Traditional sandbox solutions often struggle with the bursty nature of agentic workloads, where thousands of environments may need to be spawned and torn down rapidly. DSec’s elastic compute model is specifically engineered to absorb these fluctuations without degrading performance for active agents.

Furthermore, the system’s ability to integrate seamlessly with RL frameworks means that developers can focus on algorithmic improvements rather than infrastructure engineering. This separation of concerns is essential for accelerating innovation in agentic AI.

As the field moves toward more sophisticated autonomous agents capable of long-horizon reasoning and complex tool use, infrastructure like DSec will become increasingly vital. The ability to efficiently manage millions of concurrent sandbox environments directly impacts the feasibility of training larger, more capable models.

The open publication of the DSec architecture on arXiv invites the broader research community to build upon these foundations. With detailed documentation of the placement strategies, memory management techniques, and filesystem integrations, other organizations can potentially adopt similar approaches to meet their own scaling challenges.