kvcache-ai/AgentENV

What it solves

AgentENV provides a scalable platform for running the isolated environments (sandboxes) that AI agents need to execute code or interact with software. It specifically addresses the challenges of running millions of diverse environments across a cluster without exhausting local disk space or wasting CPU and memory on idle sandboxes.

How it works

AgentENV uses Firecracker microVMs to create lightweight, isolated sandboxes. It employs several technical optimizations to maintain high density and speed:

• Image Management: Uses overlaybd to load OCI-compatible images on demand, allowing the total number of images to exceed local disk capacity.
• Fast Lifecycle: Implements snapshot-backed booting and resuming (under 50ms) and pausing (under 100ms) to minimize resource waste.
• State Management: Supports incremental memory and filesystem snapshots and the ability to "fork" a running environment into multiple independent sandboxes.
• Resource Efficiency: Uses ublk for high-performance I/O and memory ballooning to reclaim guest memory, enabling high memory overcommit ratios.

Who it’s for

It is designed for developers and researchers training agentic RL (Reinforcement Learning) models—such as Kimi K3—and those who need to deploy large-scale, E2B-compatible agent sandboxes.

Highlights

• Massive Scale: Capable of scaling to 1.5 million images in production.
• Rapid Recovery: Boots or resumes environments in under 50ms.
• Forking Support: Allows a running environment to be cloned into multiple parallel workflows.
• E2B Compatibility: Exposes an HTTP API compatible with E2B SDKs, allowing for easy integration with existing agent tools.