AgentENV (AENV) is a distributed platform for running agent environments at scale. It gives AI teams a way to spin up, pause, snapshot and fork massive numbers of sandboxed environments across a cluster, and it powers agentic reinforcement-learning training for Moonshot AI’s Kimi models.
AgentENV is open source on GitHub with about 3,670 stars and active development in 2026. It runs Firecracker microVM environments across machines, loading diverse OCI-compatible images on demand via overlaybd and scaling to 1.5 million images in production use. Snapshot-backed environments boot or resume in under 50 milliseconds and pause in under 100 milliseconds, so idle environments release CPU and memory almost immediately. Snapshots capture memory and filesystem changes incrementally, persist to S3-compatible storage, and let a running environment fork into multiple independent sandboxes for parallel agent workflows. A local disk cache keeps hot images close to the machines that run them while cold data is evicted automatically, so cluster-wide startup stays fast even when the aggregate image and snapshot footprint exceeds local storage by several orders of magnitude. Full documentation is published online, covering cluster setup, image management and the snapshot APIs in detail.
AgentENV is free and open source under the MIT license. You run it on your own cluster; the only costs are your infrastructure and object storage.
AI platform and research teams that train or evaluate agents and need fast, cheap, massively parallel sandboxed environments with snapshot and fork semantics.