DeepSeek's Big Backend Push: 150 Infrastructure Engineers to Scale Agent Sandboxes

DeepSeek is staffing up aggressively on the infrastructure side. The company announced roughly 150 engineering positions focused on server-side systems and agent computing infrastructure—not model research—as it gears up to handle massive scaling challenges ahead.

According to Cui Tianyi, a member of DeepSeek's Harness team, the hiring concentrates on the unglamorous but critical layers: operating systems, virtualization, networking, storage, and resource coordination. The work reflects a deeper truth about where AI companies' biggest headaches lie once agents move beyond demos and into production.

The Sandbox Challenge#

Running AI agents requires far more than GPU inference. Each agent needs an isolated sandbox where it can execute code, invoke tools, modify files, and capture results safely. DeepSeek's answer is DSec (DeepSeek Elastic Compute), a sandbox infrastructure the company built to execute agent workloads during training and evaluation.

DSec supports four sandbox environments through a single Python SDK:

• Pre-warmed containers for simple function calls
• Docker-compatible containers for persistent environments
• Firecracker microVMs for stronger isolation
• QEMU virtual machines for full guest OS workloads

This versatility means the same infrastructure handles everything from a quick tool invocation to a complex software-engineering task requiring an entire operating system—a problem many AI labs are now confronting. OpenAI recently designed custom inference silicon specifically to handle agent compute pressure, while DeepSeek open-sourced its own agent harness in August.

Defeating Storage and Memory Overhead at Scale#

Copying full container or VM images to every host would be ruinous: massive storage footprint, network congestion, and slow startup times. DeepSeek's solution leans on 3FS, a distributed filesystem originally built for its training infrastructure. Container base images and filesystem commits become read-only layers backed by 3FS, with metadata kept local and data blocks fetched only when accessed—a lazy-loading approach that avoids redundancy.

The same pattern applies to microVMs: they share a read-only base layer through 3FS while individual sandbox changes stay in local copy-on-write layers. These optimizations reduce duplicate caching across virtualized environments and allow safe memory overcommitment. DeepSeek also reworked the container runtime to cut CPU overhead per sandbox—a critical tweak because spinlock contention inside the runtime, barely noticeable at small scale, becomes a density bottleneck when you're packing hundreds of thousands of sandboxes onto a single cluster.