DeepSeek describes DSec, the sandbox platform that runs 3 million agent training environments a day on 160 nodes

One production unit of DeepSeek's agent training platform starts more than 5,000 sandboxes every second, and keeps over 380,000 of them open at once. On 19 September 2026 the lab posted a 31 page technical report on arXiv describing DeepSeek Elastic Compute, or DSec, the sandbox platform behind its agent training and evaluation runs.

The four sandbox kinds behind one SDK

DSec exposes four sandbox kinds through one SDK: a plain function call, a container, a microVM, and a full VM. The abstract says one production unit spans about 160 nodes, serves around 3 million sandboxes per day, keeps over 380,000 concurrent sandboxes running, and sustains over 5,000 sandbox creations per second. Sandboxes hold state across long agent runs, and images load on demand from Fire-Flyer File System, or 3FS, DeepSeek's own cluster-wide filesystem.

Why the sandbox has to outlive its trainer

The report says DSec is co-designed with the reinforcement learning code, and that it "decouples stateful rollout execution from preemptible GPU training". In plain terms, the sandbox where the agent takes actions stays alive when the trainer next to it is paused and its GPUs are given to another job. The paper also says the platform combines memory sharing, reclamation, and CPU scheduling for high-density execution, and includes a step for what it calls "mitigating agent misbehavior such as reward hacking". Every one of those is a claim DeepSeek makes about its own system, and the paper is the place to read the method.

The gaps in the abstract

The abstract on arXiv does not give a per-sandbox cost, a hardware bill of materials, or a comparison to any named alternative. The 3 million per day figure is stated for one production unit, and the paper does not say how many units DeepSeek runs in total. The arXiv comment records that an earlier two page abstract went to first-round review for the ACM SIGOPS ATC 2026 systems track, so the peer review status is partial. DSec itself is described in the paper, not shipped as an open-source project.

For a team building a coding agent, the interesting part is the shape of the numbers. An agent that runs one tool call per second against a sandbox needs an execution layer that survives crashes, holds state for the length of a rollout, and starts a new sandbox in under a second. DSec is written up as one working answer to that question, at a scale that is likely to reset what a serious agent training or evals shop plans for.

Source

• DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale, arXiv 2609.22978, 19 September 2026
• Story surfaced on Lobsters on 23 September 2026