Kimi AI and kvcache-ai Open Source ‘AgentENV’: A Distributed System that Powers Agentic Reinforcement Learning (RL) Training for Kimi K3

Moonshot AI’s Kimi team and kvcache-ai open-sourced AgentENV (AENV), a distributed environment platform for agentic reinforcement learning.

Built on Firecracker microVMs, snapshotting, forking, and shared storage, it enables fast boot, pause/resume, and parallel rollouts.

The system was used for Kimi K3 training and is designed to make large-scale agent RL more efficient and practical.

It tackles the core bottleneck of isolated, reproducible computer environments by offering better speed than full VMs and stronger security than containers.

Released under the MIT license, AgentENV can integrate with Axum, envd, overlaybd, ublk, iroh, and S3-compatible object storage.