Powering the 2.8-trillion-parameter Kimi K3, Moonshot AI’s AgentENV provides Firecracker microVM sandboxing to execute parallel, scale-out agent workloads under an MIT license.
On 27 July, 2026, Moonshot AI and kvcache-ai officially open-sourced AgentENV alongside the model weights and training infrastructure for Kimi K3. Moonshot AI first introduced its flagship Kimi K3 model and its supporting agentic infrastructure internally and via early preview access on 16–17 July, 2026.
AgentENV (AENV) is an open-source project hosted by kvcache-ai and distributed under the permissive MIT License. The software acts as a high-throughput, distributed runtime platform that runs isolated Linux sandboxes for training AI agents via reinforcement learning (RL). AgentENV was engineered specifically to power the agentic reinforcement learning training of Kimi K3, Moonshot AI’s 2.8-trillion-parameter Mixture-of-Experts (MoE) model.
Architecturally, the platform utilizes Firecracker microVMs to grant each sandbox an independent Linux kernel, network namespace, and dedicated filesystem for strong security isolation. Root filesystems are served through a ublk userspace block device backed by overlaybd layered images, enabling shared read-only base layers and copy-on-write upper layers. An internal daemon named envd operates inside each guest VM on port 49983 to manage command execution, file operations, and health checks.
In terms of performance, the project reports boot and resume times of under 50 milliseconds, pause latencies under 100 milliseconds, and incremental snapshot captures under 100 milliseconds even during active disk writes. A single running sandbox can be cloned into up to 16 independent child sandboxes on the same node, allowing parallel exploration of task rollouts during RL loops.
For ecosystem integration, AgentENV exposes an HTTP API compatible with the E2B SDK, allowing developers to run existing Python or TypeScript agent code against it without modification by updating the API endpoint URL. The platform requires Linux Kernel 6.8 or newer with /dev/kvm access and supports single-node systemd scripts, Docker containers, Docker Compose, and Kubernetes DaemonSet clusters.
