agentenv

AgentENV (AENV) is a distributed platform designed for running agent environments at scale, notably powering agentic RL training for Kimi K3. It enables massive deployment of Firecracker microVMs across machines, loading diverse OCI-compatible images on demand via overlaybd. The system uses local disk as a bounded cache, retaining hot data and evicting cold, allowing aggregate image and snapshot footprints to exceed local disk capacity by orders of magnitude while maintaining fast cluster-wide startup without pre-warming every host. Key capabilities include: - **Scale across diverse environments**: Runs massive numbers of Firecracker environments, scaling to 1.5 million images in production. - **Inexpensive idle environments**: Snapshot-backed environments boot or resume in under 50 ms and pause in under 100 ms, allowing idle environments to quickly release CPU and memory. - **Native snapshot and fork support**: Snapshots memory and filesystem changes incrementally in under 100 ms; running environments can fork into multiple independent sandboxes for parallel agent workflows. Snapshots persist to S3-compatible object storage or shared distributed filesystems. - **Performance and density preservation**: Delivers high-performance I/O via ublk while sharing the host page cache across storage and memory-snapshot data. Memory ballooning returns reclaimable guest memory to the host, achieving a 9.6x memory overcommit ratio in production. Prerequisites: Linux kernel 6.8+ and /dev/kvm access for Firecracker microVM execution. For servers without standard KVM, a PVM deployment guide is available. Quick start (single node) involves installing the server (via install script or Docker), installing the aenv CLI, authenticating with the generated API key, pulling a template (e.g., ubuntu:22.04), and starting a sandbox. Deployment supports Docker Compose, Kubernetes clusters, and build-from-source. AgentENV exposes an E2B-compatible HTTP API, allowing use of standard E2B Python/TypeScript SDKs without code changes. The aenv CLI provides commands for template management (pull, list), sandbox lifecycle (start, pause, resume, delete, exec), and reattachment (cn). The project welcomes contributions via bug reports, feature proposals, documentation fixes, and pull requests, following CONTRIBUTING.md and SECURITY.md guidelines.