Yingdi Shan
Assistant Researcher, Department of Computer Science and Technology, Tsinghua University; Founder of the Open-Source Project AgentENV
Yingdi Shan is an Assistant Researcher in the Department of Computer Science and Technology at Tsinghua University and the founder of the open-source project AgentENV. He received his Ph.D. from Tsinghua University, with research interests focused on distributed systems and storage systems. His research has been published at leading international conferences, including OSDI, SOSP, USENIX ATC, and SIGMOD, and has received awards such as the Huawei Global Olympus Award. He has also led or participated in multiple national key R&D programs and industry collaboration projects.
Topic
AgentENV: The Large-Scale Agent Execution Environment Behind Kimi K3
AgentENV (AENV) is a distributed execution environment designed for AI Agents, jointly developed by the KVCache.AI team at the Department of Computer Science and Technology, Tsinghua University, and Moonshot AI. The project is now open source on GitHub. AI Agents need to perform complex operations in real software environments during reinforcement learning, model evaluation, and online serving, including reading code, installing dependencies, and interacting with databases. Existing sandbox platforms, however, often face challenges in cost and scalability. AENV addresses these challenges with several innovations. On-demand image pulling and layered block storage enable highly scalable execution environments; memory reuse, shared page caching, and Balloon technology significantly increase deployment density on a single machine; and suspendable snapshots substantially improve resource utilization. AENV is currently deeply integrated into the reinforcement learning training of Kimi K3. In production, it has successfully supported concurrent calls involving more than 1.5 million images and 50 million execution environments, significantly reducing infrastructure costs while providing a scalable foundation for large-scale Agent workloads.
