DeepSeek can already run hundreds of thousands of AI agent sandboxes at the same time on a single cluster, and now the company is hiring aggressively to manage what comes next as that volume continues to climb. Cui Tianyi, who joined DeepSeek in March and works on its Harness team, announced roughly 150 engineering openings on September 7, with the positions focused on server-side engineering and Agent Elastic Compute rather than AI research. The roles span operating systems, virtualization, networking, storage, scheduling, and the control-plane services that coordinate those resources, according to an X post from Cui.
At the heart of that scaling effort is DeepSeek Elastic Compute, or DSec, the sandbox infrastructure the company built to execute agent workloads during post-training and evaluation. DSec supports four types of isolated environments through the same Python SDK: pre-warmed containers for simple function calls, Docker-compatible containers for jobs requiring a persistent environment, Firecracker microVMs when stronger isolation is necessary, and QEMU virtual machines for workloads that need a full guest operating system. To avoid copying complete container or VM images onto every host—which would consume massive storage and network bandwidth while slowing startup times—DeepSeek connects DSec to 3FS, the distributed filesystem it originally created for its AI infrastructure, and stores container base images and filesystem commits as read-only layers backed by 3FS. The metadata stays local, but the underlying data blocks are retrieved only when actually required, while microVMs share their read-only base layer through 3FS as writes from individual sandboxes remain in local copy-on-write layers.
The report states that DSec reduces duplicate page-cache usage across virtualized environments and reclaims memory to permit safe overcommitment, while modifications to the container runtime lower the CPU overhead of each sandbox. During reinforcement learning and other post-training workloads, large numbers of agent rollouts can run simultaneously, and jobs may be interrupted as compute gets reassigned, but starting over wastes everything the agent has already accomplished. DSec solves this with a globally ordered trajectory log that records commands along with their results, allowing rollouts to fast-forward through completed work using those recorded results rather than executing the commands a second time. According to the report, DeepSeek said in June that it planned to at least double the size of every department, but this hiring round leans heavily toward the systems underneath its models rather than the models themselves.
Agent workloads create more than just inference demands—each agent also needs an isolated environment to execute code, call tools, modify files, and gather results, which creates compute pressure across the entire stack. The team had to address spinlock contention inside the container runtime, which barely registers at small scale but limits how densely those environments can be packed onto each host at scale. The same infrastructure can handle anything from a simple tool call to a software-engineering task that requires an entire operating system, which mirrors the challenge the rest of the industry faces as agents move from demos to production—OpenAI, for instance, recently designed custom silicon specifically to address the compute pressure that agent workloads create. Cui said DeepSeek's existing backend systems will need upgrades, maintenance, and rewrites as workloads grow, with hundreds of thousands of sandboxes running concurrently and putting pressure on everything from how jobs are scheduled to how they recover after an interruption. The company's decision to concentrate hiring in infrastructure engineering rather than research signals where the bottlenecks are forming as AI systems scale beyond training into deployment. For companies building agentic systems, the engineering required to coordinate compute at this density may prove just as strategically important as the models themselves.