DeepSeek is hiring 150 engineers, and none of them will touch a model
The New Stack Amanda Caswell
DeepSeek is hiring about 150 engineers for backend work, not model training. It’s a sign agent infrastructure is becoming the real bottleneck.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek is adding roughly 150 engineering jobs, and the twist is that none of them are for model research. The company is staffing for the plumbing underneath its AI agents: server-side systems, Agent Elastic Compute, and the infrastructure that runs and evaluates sandboxes.
That matters because DeepSeek says its clusters are already running hundreds of thousands of AI agent sandboxes at once. As those numbers climb, so do the demands on training, evaluation, scheduling, storage, networking, and the control-plane services that keep all of it from falling apart. Cui Tianyi, who works on DeepSeek’s Harness team and joined in March, said the existing backend will need upgrades, maintenance, and rewrites as workloads keep growing.
The center of that effort is DeepSeek Elastic Compute, or DSec, the sandbox system DeepSeek built for post-training and evaluation. It supports four kinds of environments through a single Python SDK: pre-warmed containers for simple function calls, Docker-compatible containers for persistent jobs, Firecracker microVMs for stronger isolation, and QEMU virtual machines when an agent needs a full guest operating system. That range lets the same stack handle everything from a quick tool call to a software task that needs an entire OS.
DeepSeek is also trying to keep all those sandboxes from crushing storage and startup times. Instead of copying whole images everywhere, DSec ties into 3FS, the distributed filesystem DeepSeek built for its AI infrastructure, and uses read-only layers with metadata kept local. The underlying blocks are fetched only when needed. For microVMs, the read-only base layer is shared through 3FS while writes stay in local copy-on-write layers.
There’s also the messy part: interruptions. During reinforcement learning and other post-training work, jobs get paused and compute gets reassigned, which means a simple restart can waste all the work an agent already did. DSec handles that with a globally ordered trajectory log that records commands and results, so a rollout can resume without replaying broken history and drifting into a different state. DeepSeek says the system also cuts duplicate page-cache use, reclaims memory for safer overcommitment, and reduces sandbox CPU overhead by dealing with spinlock contention in the container runtime.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI nobody likes to headline: not the model, but the machinery around it. DeepSeek hiring for sandboxes, schedulers, and recovery logs says agents are turning into an infrastructure business fast. The researchers get the glory; the systems engineers get the bill and the pager.
Read more about this at: The New Stack
