AgentENV

Tracked

Distributed platform for running agent environments at scale, powering agentic RL training for Kimi K3.

Author kvcache-ai Open Sourced 2026-07-23 Last Commit Unknown

Overview

AgentENV (AENV) is an open-source distributed platform for running agent environments at scale. It runs large numbers of Firecracker microVM environments across machines using diverse OCI-compatible images, powering agentic RL training for Kimi K3.

Key Features

  • Scale across environments: Runs massive Firecracker environments cluster-wide with on-demand OCI images via overlaybd; local disk acts as a bounded cache, evicting cold data so images can exceed disk capacity without pre-warming hosts
  • Inexpensive idle environments: Snapshot-backed environments boot or resume in under 50 ms and pause in under 100 ms; idle environments release CPU and memory, returning when new work arrives
  • Native snapshot and fork: Increments memory and filesystem snapshots in under 100 ms under heavy disk modification; a running environment can fork into multiple independent sandboxes for parallel agent workflows
  • High density over time: High-performance I/O via ublk with shared host page cache; memory ballooning returns reclaimable guest memory to sustain high overcommit
  • E2B-compatible API: Point E2B_API_URL at the server and use the standard E2B Python / TypeScript SDK without code changes

Use Cases

  • Agentic reinforcement learning training at scale
  • Parallel agent workflows via environment forking
  • High-density sandboxed code execution for coding agents
  • Long-running agent environments with snapshot persistence

Technical Details

  • Built on Firecracker microVMs requiring Linux kernel 6.8+ and /dev/kvm access
  • Snapshot persistence to S3-compatible object storage or shared distributed filesystem
  • Ships aenv CLI and server; deployable via systemd, Docker, Docker Compose, or Kubernetes