Kimi 团队开源 AgentENV：为 2.8 万亿参数模型提供智能体强化学习训练

Moonshot AI 的 Kimi 团队和 kvcache-ai：https://github.com/kvcache-ai 已开源了 AgentENV (AENV)：https://github.com/kvcache-ai/AgentENV，这是一个用于大规模运行代理环境的分布式平台。AgentENV 支持 Kimi K3 的代理强化学习（RL）训练，Kimi K3 是 Moonshot 的 2.8 万亿参数专家混合模型（Mixture-of-Experts）。该代码采用 MIT 许可证发布。

代理强化学习不仅仅是生成文本。它要求模型在真实计算机内执行操作。每一次回滚都需要一个带有文件系统、网络堆栈和活动进程的隔离 Linux 环境。

这一要求带来了一个难以权衡的选择。容器启动快，但共享主机内核，这削弱了对模型生成代码的隔离。完整虚拟机可以正确隔离，但启动慢，空闲时仍占用内存。

AgentENV 正是针对这一空缺设计的。它运行 Firecracker：https://github.com/firecracker-microvm/firecracker 微型虚拟机，然后使空闲、重启和分支操作足够便宜，从而可以在训练规模下运行。

每个沙箱都是一个带有独立 Linux 内核、文件系统和网络命名空间的 Firecracker 微型虚拟机。请求通过 Axum HTTP API 接入，然后转发给管理沙箱生命周期的调度器。

存储设计非常有趣。根文件系统通过由 overlaybd：https://github.com/containerd/overlaybd 分层镜像支持的 ublk 用户空间块设备提供。只读基础层在沙箱间共享，每个沙箱将数据写入其自己的上层。

在每个客户机内，一个名为 envd 的守护进程负责命令执行、文件操作和在端口 49983 上的健康报告。反向代理将客户端的 HTTP 和 WebSocket 流量路由到 VM 内运行的服务。

该项目还列出了两种密度机制。主机页缓存跨存储和内存快照数据共享。内存气球技术将可回收的客户机内存返回给主机，从而在环境随时间变化时维持过量分配。

这四个功能是该项目存在的原因。AgentENV 以增量方式快照内存和文件系统的变化，而不是每次都写入完整镜像。

报告的数字如下。以快照为基础的环境启动或恢复时间不到50毫秒，暂停时间不到100毫秒。在磁盘修改繁重的情况下，增量快照捕获也能在100毫秒内完成。

Fork 是 RL 最独特的功能。运行中的沙箱可以在同一节点上克隆成最多16个独立的子沙箱。源沙箱在捕获期间会短暂暂停，然后恢复。每个子沙箱继承源沙箱的文件系统、内存和资源配置。

实际效果是昂贵的设置只需运行一次。团队可以安装依赖项、克隆代码库并达到任务状态。然后，该确切状态可以分支为并行部署。快照会持久化到兼容 S3 的对象存储或共享分布式文件系统中。

有一个默认值值得注意。每个沙箱都有 TTL，过期会触发暂停，而非删除。删除需要在创建 API 中传递 autoPause: false。

镜像通过 overlaybd 按需加载。本地磁盘充当有界缓存，保留热数据并驱逐冷数据。

这就是使集群级方案可行的原因。节点不需要预热每个镜像或保留每个快照的完整副本。因此，可寻址的镜像集合可以超过本地磁盘容量，同时启动在整个集群中保持快速。

快照状态分为三层。构建器暂存工作区在构建过程中保存工件。已提交的快照仓库是持久化的真实来源。节点本地运行时缓存保存启动时派生的配置。

支持两种仓库后端：posix_fs（默认）和 oss。oss 路径通过共享的 S3 兼容客户端运行，因此需要指定区域。

可选的基于 iroh 的点对点传输可以向同伴节点公布已提交的工件。默认情况下禁用。文档明确指出 P2P 不会改变已提交的快照模型。对于共享存储，文档要求至少 1 Gbps，并强烈推荐 10 Gbps 或更快。

AgentENV 提供 E2B 兼容的：https://kvcache-ai.github.io/AgentENV/integration/e2b.html HTTP API。将 E2B_API_URL 指向你的服务器，官方 E2B：https://github.com/e2b-dev/e2b Python 或 TypeScript SDK 可在不更改代码的情况下使用。

这是一个有意的分发选择。已经在 E2B 上运行代理的团队可以自托管运行时，而无需重写他们的代理代码。还提供了一个原生的 aenv CLI，文档推荐在 AgentENV 特定工作流程中使用它。

先决条件是 Linux 内核 6.8 及以上以及 /dev/kvm 访问；安装脚本还需要 Ubuntu 24.04。aenv CLI 支持 x86_64 和 arm64 架构的 Linux 和 macOS。服务器本身仅支持 Linux，因为它需要 KVM。

多节点部署在 :8080 添加一个网关，在 :9090 添加一个调度器。

请查看 GitHub 仓库：https://github.com/kvcache-ai/AgentENV 以及文档：https://kvcache-ai.github.io/AgentENV/。所有研究成果归此项目研究人员所有。

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一名富有远见的企业家和工程师，Asif 致力于利用人工智能的潜力造福社会。他最近的努力是推出人工智能媒体平台 Marktechpost，该平台因其对机器学习和深度学习新闻的深入报道而脱颖而出，这些报道既技术可靠，又易于广大读者理解。该平台月访问量超过 200 万，显示出其在观众中的受欢迎程度。

构建一个代理事件场馆运营商 [完整代码]：https://pxllnk.co/twdn5

谢谢！我们的团队将很快与您联系 🙌

Moonshot AI’s Kimi team and kvcache-ai：https://github.com/kvcache-ai have open-sourced AgentENV (AENV)：https://github.com/kvcache-ai/AgentENV , a distributed platform for running agent environments at scale. AgentENV powers agentic reinforcement learning (RL) training for Kimi K3 , Moonshot’s 2.8-trillion-parameter Mixture-of-Experts model. The code ships under an MIT license .

Agentic RL does not just sample text. It requires the model to act inside a real computer. Every rollout needs an isolated Linux environment with a filesystem, a network stack, and live processes.

That requirement creates a hard trade-off. Containers start fast but share the host kernel, which weakens isolation for model-generated code. Full virtual machines isolate properly but boot slowly and hold memory while idle.

AgentENV targets exactly that gap. It runs Firecracker：https://github.com/firecracker-microvm/firecracker microVMs, then makes idle, restart, and branching cheap enough to run at training scale.

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

Storage is where the design gets interesting. The rootfs is served through a ublk userspace block device backed by overlaybd：https://github.com/containerd/overlaybd layered images. Read-only base layers are shared across sandboxes, and each sandbox writes into its own upper layer.

Inside every guest, a daemon called envd handles command execution, file operations, and health reporting on port 49983 . A reverse proxy routes HTTP and WebSocket traffic from clients to services running inside the VM.

The project also lists two density mechanisms. The host page cache is shared across storage and memory-snapshot data. Memory ballooning returns reclaimable guest memory to the host, which sustains overcommit as environments diverge over time.

These four features are the reason the project exists. AgentENV snapshots memory and filesystem changes incrementally rather than writing a full image each time.

The reported figures are as follows. Snapshot-backed environments boot or resume in under 50 ms and pause in under 100 ms . Incremental snapshot capture completes in under 100 ms , even under heavy disk modification.

Fork is the feature most specific to RL. A running sandbox can clone into up to 16 independent child sandboxes on the same node. The source pauses briefly during capture, then resumes. Every child inherits the source filesystem, memory, and resource configuration.

The practical effect is that expensive setup runs once. A team can install dependencies, clone a repository, and reach a task state. That exact state then branches into parallel rollouts. Snapshots persist to S3-compatible object storage or a shared distributed filesystem.

One default is worth noting. Every sandbox carries a TTL, and expiry triggers a pause , not a delete. Deletion requires passing autoPause: false on the create API.

Images load on demand through overlaybd. Local disk acts as a bounded cache that retains hot data and evicts cold data.

This is what makes the fleet-level story work. Nodes do not need to pre-warm every image or hold a complete copy of every snapshot. The addressable image set can therefore exceed local disk capacity while startup stays fast cluster-wide.

Snapshot state is organized in three layers. A builder staging workspace holds artifacts during a build. A committed snapshot repository is the durable source of truth. A node-local runtime cache holds launch-time derived configs.

Two repository backends are supported: posix_fs (default) and oss . The oss path runs through a shared S3-compatible client, so an explicit region is required.

An optional peer-to-peer transport based on iroh can advertise committed artifacts to peer nodes. It is disabled by default . The documentation is explicit that P2P does not change the committed snapshot model. For shared storage, the docs ask for at least 1 Gbps and strongly recommend 10 Gbps or faster.

AgentENV exposes an E2B-compatible：https://kvcache-ai.github.io/AgentENV/integration/e2b.html HTTP API . Point E2B_API_URL at your server and the official E2B：https://github.com/e2b-dev/e2b Python or TypeScript SDK works without code changes.

This is a deliberate distribution choice. Teams already running agents on E2B can self-host the runtime without rewriting their agent code. A native aenv CLI is also provided, and the docs recommend it for AgentENV-specific workflows.

Prerequisites are Linux kernel 6.8+ and /dev/kvm access; the install script additionally requires Ubuntu 24.04. The aenv CLI supports Linux and macOS on x86_64 and arm64. The server itself is Linux-only, because it needs KVM.

Multi-node deployments add a gateway on :8080 and a scheduler on :9090 .

Check out the GitHub repository：https://github.com/kvcache-ai/AgentENV and documentation：https://kvcache-ai.github.io/AgentENV/. All credit for this research goes to the researchers of this project.

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

Build an Agentic Event Venue Operator [Full Codes]：https://pxllnk.co/twdn5

Thanks! Our team will contact you soon 🙌