How DeepSeek trains agents on 380,000 sandboxes at once
DeepSeek's DSec paper shows the sandbox platform behind agentic RL from V3.2 to V4.1: 3M sandboxes a day, 380K concurrent, 5,000 creations per second. The key insight is that agent sandboxes are almost always idle, and the hard problems are images, memory, and agents that cheat.
DeepSeek-AI, Tsinghua University