DeepSeek is opening roughly 150 backend engineering positions to expand DSec, its sandbox platform capable of running hundreds of thousands of concurrent agent environments across a single cluster. The staffing push reflects the company's need to manage infrastructure as workload volumes continue to increase.

Cui Tianyi, who joined DeepSeek in March as part of the Harness team responsible for agent infrastructure and evaluation environments, disclosed the hiring initiative on September 7 via X. The positions concentrate on server-side engineering and Agent Elastic Compute roles rather than machine learning research, covering domains such as operating systems, virtualization, networking, storage, scheduling, and control-plane services that manage these resources.

DeepSeek's existing backend systems will need upgrades, maintenance, and rewrites as workloads grow.

Cui Tianyi

The scaling challenge centers on DeepSeek Elastic Compute, or DSec, the sandbox infrastructure built to execute agent workloads during post-training and evaluation phases.

Four sandboxes, one SDK

Running agent workloads requires more than GPU inference capacity; each agent needs an isolated execution environment for running code, invoking tools, modifying files, and capturing outputs. DSec provides four environment types through a single Python SDK: pre-warmed containers for straightforward function calls, Docker-compatible containers for jobs requiring persistent state, Firecracker microVMs for stronger isolation requirements, and QEMU virtual machines for workloads needing a complete guest OS.

This flexibility enables the same infrastructure to support everything from basic tool invocations to complex software-engineering tasks demanding full operating system access. The industry faces similar challenges as agent systems transition from experimental prototypes to operational deployments. OpenAI recently developed custom silicon to address computational demands from agent workloads, while DeepSeek released its own agent harness as open source in August.

Lazy loading agent environments

Each sandbox requires its own environment, yet distributing complete container or VM images to every host would consume substantial storage and network resources while increasing startup latency. DeepSeek resolves this by integrating DSec with 3FS, a distributed filesystem originally developed for its AI infrastructure, and storing container base images and filesystem commits as read-only layers backed by 3FS. Metadata remains local while underlying data blocks load on demand.

MicroVMs employ an analogous approach, accessing shared read-only base layers through 3FS while individual sandbox modifications reside in local copy-on-write layers. According to DeepSeek, DSec minimizes redundant page-cache usage across virtualized environments and recovers memory for safe overcommitment, with container runtime modifications reducing CPU overhead per sandbox.

The infrastructure team also addressed spinlock contention within the container runtime. While CPU time spent on this issue remains negligible at small deployments, it becomes a bottleneck at scale, constraining how tightly these environments can be consolidated on individual hosts.

DeepSeek says DSec reduces duplicate page-cache usage across virtualized environments and reclaims memory to allow safe overcommitment, while changes to the container runtime cut the CPU overhead of each sandbox.

When replay breaks training

Reinforcement learning and other post-training processes execute numerous agent rollouts simultaneously, and jobs may be interrupted when compute resources are reallocated. Restarting from scratch discards prior progress, yet resuming is complicated because earlier commands may have modified files or changed the environment, potentially producing different outcomes or corrupting the training state.

DSec addresses this through a globally ordered trajectory log capturing commands and their corresponding results. Upon rollout resumption, DSec fast-forwards through completed work using recorded results instead of re-executing commands, lowering interruption costs across thousands of training and evaluation jobs. The same logs maintain a record of sandbox modifications and enable earlier sessions to be replayed.

Engineers, not researchers, wanted

The approximately 150 positions span DeepSeek's backend infrastructure, encompassing lower-level systems supporting Agent Elastic Compute and services underlying its models and agents. While DeepSeek stated in June plans to at least double departmental headcount, this hiring wave prioritizes systems-level work beneath its models rather than model development itself. With hundreds of thousands of sandboxes operating concurrently, DSec creates pressure across job scheduling, interruption recovery, and related infrastructure components.

The roughly 150 openings reach across DeepSeek's backend, including the lower-level systems work behind Agent Elastic Compute as well as the services that support its models and agents.

Source: The New Stack