On September 23, 2026, DeepSeek dropped a new system paper that flew under many industry radars — but its implications for the AI agent race are enormous. Titled “DeepSeek Elastic Compute: Sandbox Infrastructure for Massive-Scale Agent Training” and authored by a team led by founder Liang Wenfeng, the paper introduces Dsec — a production-grade sandbox platform that now powers all of DeepSeek’s agent reinforcement learning and evaluation workloads, from V3.2 through V4.1.
This is not another model benchmark. It is the answer to a question the industry has been quietly grappling with: How do you train agents that actually interact with the world at scale?
Why Agent Training Needs a Different Kind of Infrastructure
Traditional LLM training — even with DeepThink’s GRPO reasoning loop — happens in a relatively controlled environment. The model predicts tokens, gets reward signals, and updates its weights. But when an agent needs to run code, browse the web, call APIs, or execute multi-step tasks in a terminal, the training landscape changes entirely:
- Agents need ephemeral execution environments. A single training run might spin up thousands of sandboxes, each running code, accessing files, and making network requests — then tear them down seconds later.
- Parallelism requirements explode. Instead of GPU clusters churning through tensor operations, you need CPU clusters managing container lifecycles, network isolation, and resource quotas for millions of concurrent sandbox instances.
- State management becomes critical. An agent that makes progress through a multi-step task needs its sandbox state preserved across episodes, but also needs clean-room isolation for evaluation.
Most agent frameworks today rely on ad-hoc Docker or Kubernetes setups. These work for demos and small-scale experiments — but they break down when you need to run millions of sandboxes per day as part of a production training pipeline.
Dsec: The Numbers Behind the Platform
DeepSeek’s paper reveals why Dsec is more than just an internal tool:
| Metric | Dsec Production Unit |
|---|---|
| CPU Nodes | 160 nodes |
| CPU Cores | 30,000+ cores |
| DRAM | 250 TB |
| Daily Sandbox Services | ~3 million |
| Peak Concurrent Sandboxes | 380,000+ |
| Sandbox Creation Rate | 5,000+ per second |
A single Dsec production unit serves millions of agent training episodes daily, with peak concurrency that rivals the container fleets of major cloud providers.
The Elastic Compute Advantage
The key architectural insight of Dsec is elastic compute with on-demand image loading. Rather than pre-pulling container images to every node (which wastes disk space and network bandwidth), Dsec loads only the image layers needed for each specific sandbox on demand.
In benchmarks, this approach reduced startup time for 8,192 containers to approximately 35 minutes — a 42% improvement over full remote pull strategies — while cutting disk write volume by 57%.
For agent training, this matters enormously: every second saved in sandbox startup is more time spent on actual reinforcement learning iterations. And with DeepThink’s reasoning loop requiring multiple agent attempts per training episode, those savings compound quickly.
How Dsec Enables DeepThink’s Next Generation
DeepSeek’s latest V4.1 Flash model demonstrated that efficient inference at low cost is achievable. But the company’s roadmap — hinted at in their recent hiring spree for 150+ “agent elastic compute” engineers — suggests that training infrastructure is now the bottleneck for the next leap in agent capability.
Dsec solves three critical constraints:
1. Scaling Agentic RL Without Breaking the Bank
Agent reinforcement learning is computationally expensive. Each episode might require an agent to make dozens of tool calls, execute code in a sandbox, and receive reward signals. Dsec’s high-density sandbox scheduling means DeepThink can train agents on more diverse environments without proportional cost increases.
2. Reproducible Evaluation at Scale
One of the industry’s biggest challenges with agent benchmarks is reproducibility. Dsec’s isolated sandboxes with deterministic resource quotas make it possible to run the same agent evaluation thousands of times with consistent conditions — essential for reliable A/B testing during training.
3. Security by Design
When agents can execute arbitrary code, security is non-negotiable. Dsec’s sandbox isolation provides a clean boundary between the agent’s execution environment and the training infrastructure, preventing reward hacking attempts that target the evaluation infrastructure itself.
The Competitive Landscape
DeepSeek is not alone in building specialized agent training infrastructure. OpenAI reportedly invested heavily in custom execution environments for their operator model, and Anthropic has developed internal sandbox systems for Claude’s tool use capabilities.
But Dsec represents a different strategy: build for scale first, then optimize for specific agent workloads. By running 380,000 concurrent sandboxes in production, DeepThink has demonstrated infrastructure capability that most competitors can only prototype in research settings.
The timing is also notable. DeepSeek’s V4.1 Flash and Dsec announcements within days of each other suggest a coordinated push to demonstrate that the company leads not just in model capability, but in the entire agent system stack — from inference efficiency to training infrastructure.
What This Means for the Industry
Three takeaways stand out:
The Agent Infrastructure Arms Race Is On
As agent benchmarks like Terminal Bench and CyberGym become more prominent, the ability to train agents at scale becomes a competitive moat. Models with comparable intelligence can diverge sharply based on the quality and scale of their training environments.
Open Questions Remain
DeepSeek’s paper focuses on system architecture and performance metrics, but leaves open some critical questions: How does Dsec handle stateful multi-step agent workflows? What are the fault tolerance mechanisms for distributed training across sandbox clusters? And perhaps most importantly — will any of this infrastructure be open-sourced, like Harness was?
DeepThink’s Vision Is Getting Clearer
From DeepThink R1’s transparent reasoning to Harness’s agent orchestration framework to Dsec’s training infrastructure, a pattern emerges: DeepSeek is building a full-stack agent platform, not just another model. The question for the industry is whether competing labs will match this vertical integration, or continue focusing on model capability alone.
The Road Ahead
DeepThink’s next generation — likely built on V4.2 or V5 — will be shaped by what Dsec enables. With millions of daily sandbox episodes and 5,000+ creation rate per second, the company can explore agent capabilities that were previously limited by infrastructure constraints: longer-horizon tasks, more diverse tool environments, and more complex multi-agent interactions.
For now, Dsec remains an internal system. But given DeepSeek’s track record of open-sourcing critical components (Harness, V4.1 Flash weights), it would not be surprising to see parts of the Dsec architecture made available to the broader developer community — further cementing DeepThink’s role as the infrastructure backbone of the 2026 agent revolution.
