wtf is agentenv?

AgentENV, also called AENV, is a platform for running huge numbers of isolated sandbox environments at once, built to support AI agents that need somewhere safe to run code and commands. It was built to power training for another AI system, Kimi K3, where thousands of these environments needed to start and stop constantly. The core idea is speed and efficiency at scale. Each sandbox is a lightweight virtual machine that can boot or resume in under fifty milliseconds and pause in under one hundred milliseconds, so idle sandboxes can quickly give back their CPU and memory instead of sitting there wasting resources. A running sandbox can also be saved as a snapshot and split into several independent copies, useful when many agent workflows need to branch off from the same starting point. These snapshots are saved to cloud storage or a shared file system so nothing is lost if a machine goes down. AgentENV can load container images on demand across many machines rather than needing every server pre loaded ahead of time, using local disk as a cache that keeps frequently used data close and discards what is not needed. In production use, this design let a deployment scale to handle over a million images while keeping startup fast everywhere. Getting started involves installing a server, either through a script or Docker, and a matching command line tool called aenv. Once running, you authenticate against the server, pull a template image such as Ubuntu, and start a sandbox with an interactive shell. AgentENV also speaks the same API as E2B, an existing sandbox service, so tools already built for E2B can point at an AgentENV server without changes. The project currently has no built in authorization, so the README is explicit that it should never be exposed to the public internet and should only run on a trusted network or behind a protected proxy. It requires a recent Linux kernel and access to KVM virtualization on the host machine.