Tencent Cloud Open-Sources Cube Sandbox for AI Agents

Tencent Cloud open-sourced its Cube Sandbox on April 21, 2026, under the Apache 2.0 license. The sandbox-as-a-service stack is designed as a production-grade foundation for deploying large-scale AI Agents, utilizing MicroVM architecture to ensure hardware-level isolation and prevent kernel escape or malicious code execution.

The system features technical capabilities including cold start times under 60ms, per-instance memory overhead below 5MB, and the capacity to burst schedule more than 100,000 instances. To facilitate adoption, Cube natively supports E2B SDKs and OpenAI Python, allowing developers to migrate without changing their code.

Dowson Tong first announced the open-source initiative on March 27 during the Tencent Cloud Shanghai City Summit. The company plans to integrate Cube with its FlexKV cache system and TACO AI Acceleration Engine to establish a complete infrastructure for secure AI Agent runtimes. The technology is already being utilized by the foundation model lab Minimax to run hundreds of thousands of concurrent heterogeneous sandboxes for Agentic RL training.