Kimi AI and kvcache-ai Open Sources ‘AgentENV’: A Distributed System that Powers Agentic Reinforcement Learning (RL) Training for Kimi K3

Moonshot AI’s Kimi staff and kvcache-ai have open-sourced AgentENV (AENV), a distributed platform for operating agent environments at scale. AgentENV powers agentic reinforcement studying (RL) coaching for Kimi K3, Moonshot’s 2.8-trillion-parameter Mixture-of-Experts mannequin. The code ships below an MIT license.

Why Environment Infra Holds Back Agentic RL

Agentic RL doesn’t simply pattern textual content. It requires the mannequin to behave inside an actual laptop. Every rollout wants an remoted Linux surroundings with a filesystem, a community stack, and dwell processes.

That requirement creates a tough trade-off. Containers begin quick however share the host kernel, which weakens isolation for model-generated code. Full digital machines isolate correctly however boot slowly and maintain reminiscence whereas idle.

AgentENV targets precisely that hole. It runs Firecracker microVMs, then makes idle, restart, and branching low-cost sufficient to run at coaching scale.

Inside AgentENV’s Firecracker Architecture

Each sandbox is a Firecracker microVM with its personal Linux kernel, filesystem, and community namespace. Requests hit an Axum HTTP API, which forwards to an orchestrator that manages the sandbox lifecycle.

Storage is the place the design will get attention-grabbing. The rootfs is served by way of a ublk userspace block machine backed by overlaybd layered pictures. Read-only base layers are shared throughout sandboxes, and every sandbox writes into its personal higher layer.

Inside each visitor, a daemon known as envd handles command execution, file operations, and well being reporting on port 49983. A reverse proxy routes HTTP and WebSocket visitors from purchasers to companies operating contained in the VM.

The undertaking additionally lists two density mechanisms. The host web page cache is shared throughout storage and memory-snapshot information. Memory ballooning returns reclaimable visitor reminiscence to the host, which sustains overcommit as environments diverge over time.

Key Takeaways

• Deployment covers set up script, Docker, Docker Compose, and Kubernetes; the multi-node management airplane is documented as a prototype.
• AgentENV runs every agent surroundings as a Firecracker microVM, not a container, so isolation is kernel-level.
• The undertaking stories boot or resume below 50 ms, and pause below 100 ms, from snapshots.
• A operating sandbox can fork into as much as 16 unbiased kids on the identical node.
• The HTTP API is E2B-compatible, so present E2B Python and TypeScript SDK code runs unchanged.

Check out the GitHub repository and documentation. All credit score for this analysis goes to the researchers of this undertaking.

The put up Kimi AI and kvcache-ai Open Sources ‘AgentENV’: A Distributed System that Powers Agentic Reinforcement Learning (RL) Training for Kimi K3 appeared first on MarkTechPost.