AgentENV is a distributed platform for running agent environments at scale using Firecracker microVMs.
The platform addresses the challenge of executing massive numbers of isolated agent environments efficiently across clusters. It uses Firecracker microVMs loaded with OCI-compatible images on demand via overlaybd, with local disk acting as a bounded cache that can retain hot data while evicting cold data. This approach allows the aggregate image and snapshot footprint to exceed local disk capacity by several orders of magnitude while maintaining fast startup times cluster-wide. Idle environments can pause in under 100 ms and resume in under 50 ms, quickly releasing CPU and memory when not in use. Memory ballooning returns reclaimable guest memory to the host, enabling high memory overcommit ratios as environments run longer.
AgentENV suits teams running agentic reinforcement learning workloads that require scaling to large numbers of isolated environments while maintaining cost efficiency during idle periods. It is designed for production deployments where performance density matters and environments need to persist state reliably. The platform requires Linux kernel 6.8 or later with KVM access, though alternative deployment guidance exists for systems without standard KVM support. The tool provides both a server component and a CLI for interaction, with setup options including an install script or Docker deployment.
The project maintains comprehensive documentation split between stable and development versions. Code coverage is tracked and reported through automated workflows. The codebase is written in Rust, indicating a focus on systems-level performance and safety.
