DeepSeek has forged a partnership with a research team from Tsinghua University to unveil DSec—an open-source sandbox infrastructure tailored for large-scale agent training. DSec streamlines the process of creating and managing sandboxes by incorporating three core features: modular layers, dynamic image loading, and precise resource allocation. In Section 6 of their paper, the researchers introduce an innovative concept: an environment crafted by agents, for agents, and owned by agents. This framework empowers agents to autonomously design training environments and generate incremental snapshots for future use. This marks the initial formation of an RSI closed loop, overcoming the previous inefficiencies associated with manually constructing reinforcement learning environments for agents.

Presently, the adoption of similar agent sandbox technologies is gaining momentum across the industry, with sandboxes evolving into a novel runtime environment for agents. The competitive landscape has shifted its focus from model performance to the underlying agent environment infrastructure, with sandbox isolation security emerging as a pivotal competitive advantage. The race to implement RSI through sandboxes has officially commenced.