deeplethe/forkd is an isolated microVM runtime for AI agents. The latest release is v0.5.3, but the provided material does not list what changed in this version. According to the repository description, it starts 100 children in roughly 100 milliseconds from a prepared parent VM, and branches a running VM in about 150 milliseconds. The project is written in Rust and follows the Apache-2.0 license.
forkd is based on Firecracker and creates child environments by copying the memory snapshot of a prepared VM and sharing it in copy-on-write (CoW) fashion. The README also describes the ability to branch a running environment and differential snapshot linking. VMs are isolated with KVM. The repository also contains a command-line tool, a programming interface, design documents, and performance experiment material. Its core purpose is to create a separate execution environment for each agent and to start child environments quickly from that state.
That said, the roughly 100 millisecond and roughly 150 millisecond figures are values stated in the repository description. The provided material contains no measurement conditions or comparison results backing them, so the same speed cannot be asserted in other environments. The v0.5.3 changelog is also unconfirmed. This repository can therefore be seen as an agent isolation runtime that promotes fast branching, but concrete performance and the improvements in this release require additional material.