Google 开源 AX：Agent 工作负载编排运行时
Google 开源 AX，声明式编排海量自主 Agent 任务的运行时，沙箱执行、检查点暂停恢复，用法类 kubectl。
来源摘录
Google just open-sourced something absolutely insane. It's called AX, and it's designed to orchestrate BILLIONS of autonomous AI agent tasks. Think Kubernetes but built for agents instead of apps. Here's why this matters: Agents are a brand new kind of workload. They're not microservices. They're not batch jobs. They hold state, need strict isolation, call models and tools nonstop, and can burn through your money in a loop if nobody's watching How it works: - You write one YAML file. Declare the task, the workspace, and the model - AX spins up a sandbox, wires in your Git repos, MCP servers, and skills so every agent starts warm - You run "ax apply" and the agent goes to work - "ax ssh" lets you shell into a running agent and watch it think - "ax suspend" pauses an idle agent. "ax resume" picks up exactly where it left off If you've ever used kubectl, you already know how to use this 13.5K stars already. Apache 2.0. Free forever Here's what I'd do this weekend: 1. Spin up a small Kubernetes cluster 2. Install Agent Substrate (AX runs on top of it) 3. Run the demo script so you see the full lifecycle end to end 4. Point one agent at a real repo and give it a real goal It's early. Google says breaking changes are coming before a stable release. Doesn't matter The people who learn agent infrastructure now are the ones who'll run 1,000 agents while everyone else is still babysitting one https://t.co/apru4DsXvM
深度解析
是什么：google/ax 是 Google 开源的声明式 Agent 编排运行时（Go 编写、Apache 2.0、约 13.9K stars），定位类似"面向 Agent 的 Kubernetes"，运行在 Agent Substrate 之上，以沙箱化方式在集群中大规模运行自主 Agent 任务。
核心设计：提供三个原语——Task（以 CPU/内存限制的沙箱运行不可信 Agent 代码）、Workspace（预接线 Git 仓库、MCP 服务器与技能包，让 Agent 热启动）、Model（配置平台自身 LLM，凭据走 K8s Secret）。一切以 ax.io/v1alpha1 YAML 声明；CLI 刻意做成 kubectl 形状：apply/get/describe/watch/ssh/suspend/resume，支持检查点暂停后原位恢复，并可用 ax ssh 进入沙箱观察 Agent 行为；控制面经 gRPC 通信，依赖 Redis，经 ko 构建部署到 ax-system 命名空间，且跟随 kubectx 上下文自动切换集群隧道。
关键洞察：Agent 是全新负载类型——既非无状态微服务也非一次性批任务，会积累状态、需强隔离、持续调用模型与工具 API、无人监管时可能陷入烧钱循环；把 K8s 成熟的声明式运维心智模型迁移到 Agent 基础设施，是学习成本最低的路径，这正是项目最值得借鉴的架构判断。
边界与风险：项目处于重度开发期，官方警告稳定版前会有重大破坏性变更；前置依赖 K8s 集群、Agent Substrate、Go/ko 与容器镜像仓库，部署链路门槛不低；推文中"数十亿任务""13.5K stars"等表述带宣传语气，具体应以 README 与仓库元数据为准。