Kimi AI Open Sources AgentENV for Agent Training

Moonshot AI's Kimi team and kvcache-ai have released AgentENV, also called AENV, as open source under an MIT license. The system handles distributed agent environments at scale and currently powers agentic reinforcement learning training for Kimi K3, Moonshot's Mixture-of-Experts model with 2.8 trillion parameters.

The release addresses a specific bottleneck in agentic RL: giving a model a real computer to act in, over and over, at training volume. Unlike standard text sampling, this kind of training needs isolated Linux environments complete with filesystems, network stacks, and live processes for every single rollout. Engineers building these systems have faced a persistent trade-off.

Containers spin up quickly but leak isolation because they share the host kernel, a problem when the code running inside was generated by the model itself. Full virtual machines solve the isolation problem but boot slowly and eat memory even when sitting idle.

AgentENV was built to close that gap using Firecracker microVMs, with the goal of making idle time, restarts, and branching cheap enough to survive the pace of large-scale training runs. The architecture behind that claim, and how storage and networking are handled inside each sandbox, is where the project gets specific.