@cloudflare/computer：给每个 Agent 一台自己的电脑

by Mycelium Protocol

---

给 Agent 开一个容器（Container），这是过去一年里大多数团队的默认答案：容器里装好文件系统、Shell、依赖包，Agent 在里面干活，干完销毁。

Cloudflare 在刚结束的 Agents Week（2026 年 8 月 2-4 日）里指出了这条路的问题：全球所有超大规模云加起来，也没有足够的计算资源给每个用户的每个 Agent 各开一个容器。1 亿并发 Agent，1 亿个容器——这不可能。

他们给出的答案是 @cloudflare/computer（早期预览，已开源）。

---

问题根源：Agent 不需要一直用容器

容器很重，但 Agent 大多数时候做的事很轻：读写文件、检查 git 状态、解析 JSON、处理数据。这些操作根本不需要完整的 Linux userland。

Cloudflare 的观察是：绝大多数 Agent 操作可以在 Isolate（Workers）里完成，只有在需要 npm、原生二进制、或者完整 Shell 环境时才需要容器。目标是让容器调用比例低于 10%。

这个观察来自他们自己在用的 Agent 架构：Agent 逻辑运行在 Durable Object（Isolate）里，容器按需作为工具挂载上来。

---

架构：共享文件系统 + 双执行后端

@cloudflare/computer 的核心是 Workspace——一个以 SQLite 为底层存储的虚拟文件系统：

Workspace（虚拟文件系统，SQLite 持久化）
   ├── 来源：git repo、云存储 bucket、任意文件
   ├── Isolate 后端（fast, cheap）
   │   └── just-bash → 翻译成 JS → Dynamic Worker 执行
   │       文件系统通过 Worker Binding 直接访问
   └── Container 后端（full Linux）
       └── Cloudflare Containers + FUSE 挂载
           修改自动同步回 Workspace

两个后端共用同一套文件，共用同一个 exec(string, options) 接口。Agent 的 system prompt 里描述两个 backend 的特点，模型自行选择——经测试，前沿模型的选择准确率很高。

---

快速上手

npm install @cloudflare/computer

最小示例：在 @cloudflare/think Agent 里挂载一个 Workspace：

import { Think } from "@cloudflare/think";
import { Workspace } from "@cloudflare/computer";
import { createWorkersAI } from "workers-ai-provider";

export class Agent extends Think {
  override workspace = new Workspace({
    storage: this.ctx.storage,
    useThink: true,
  });

  override getModel() {
    return createWorkersAI({ binding: this.env.AI })("@cf/zai-org/glm-5.2");
  }

  override getSystemPrompt() {
    return `你是一个 bug 分诊 Agent。
用 /workspace/repo 里的代码复现 bug，定位问题，在安全的情况下
做出精准修复，然后验证。`;
  }
}

接上 Cloudflare Container 后端：

import {
  CloudflareContainerBackend,
  withWorkspaceContainer,
} from "@cloudflare/computer/backends/container";

export class Agent extends withWorkspaceContainer(Think) {
  override workspace = new Workspace({
    storage: this.ctx.storage,
    useThink: true,
    backends: [
      new CloudflareContainerBackend({
        container: () => this,
        workspace: {
          binding: "Agent",
          id: this.ctx.id.toString(),
        },
      }),
    ],
  });
}

暴露文件/git/shell 工具给模型：

import { createAITools } from "@cloudflare/computer/tools";

override getTools() {
  return {
    ...createAITools({
      workspace: this.workspace,
      shell: {
        defaultBackend: "container",
        backends: {
          container: {
            description:
              "完整 Linux 环境：npm、node、包管理器、测试运行器、原生命令都在 PATH 里。" +
              "只有任务超出文件操作范围时才用它。",
          },
        },
      },
    }),
    replyToIssue,  // 你自己的产品工具
  };
}

Agent 开始前用 API 预置环境：

async startTriage(report: { title: string; body: string; repoUrl: string }) {
  await this.workspace.fs.mkdir("/workspace", { recursive: true });
  await this.workspace.fs.writeFile(
    "/workspace/BUG_REPORT.md",
    `# ${report.title}\n\n${report.body}\n`,
  );
  await this.workspace.git.clone({
    url: report.repoUrl,
    dir: "/workspace/repo",
  });

  return this.submitMessages([{
    id: crypto.randomUUID(),
    role: "user",
    parts: [{ type: "text", text: `分诊这个 bug: ${report.title}` }],
  }]);
}

Workspace 还提供 node:fs 兼容的 wrapper，可以直接配合第三方 Node.js 库使用。

---

为什么是 Isolate 而不是容器

Cloudflare 在这个方向上押注了将近 10 年：

• 2017：推出 Cloudflare Workers（Isolate 模型）
• 2020：推出 Durable Objects（持久化 Isolate，有状态，横向无限扩展）
• 2024：Durable Objects 里嵌入 SQLite，横向扩展同时获得垂直能力
• 2025：Durable Objects 可以按需挂载容器沙箱（Container 进 GA beta）
• 2026：@cloudflare/computer 把这一套封装成统一 API

Isolate 的优势：

• 毫秒级冷启动（容器是秒级）
• 空闲时自动 hibernate（不计费）
• 理论上横向无限扩展
• 可以动态派生新的 Isolate 执行不可信代码

容器的优势：

• 完整 Linux userland
• 任意二进制、任意包管理器
• 垂直扩展能力（大内存、多核）

@cloudflare/computer 让两者在同一个工作流里并存，而不是强迫开发者选择。

---

技术细节

Isolate 后端：使用 just-bash 把 Shell 命令翻译成 JavaScript，在 Dynamic Worker 里执行。Workspace 的文件通过 Worker Binding 直接可见，不需要网络传输。

Container 后端：使用 FUSE（Filesystem in Userspace）把 Workspace 挂载到容器里。文件变更实时同步回 Workspace（SQLite），Agent 状态始终一致。

audit 和 access control：所有 exec 操作都有 gate 和 audit trail，可以精确控制 Agent 允许执行哪些操作。

---

现在能用来做什么

Cloudflare 内部已经在用这套架构做：

• 仅用 Isolate 来 build、test、deploy JavaScript 应用
• 给每个客户生成定制化文档
• 用 Web Browser 执行复杂任务

开源仓库：github.com/cloudflare/computer，含 step-by-step tutorial（examples/tutorial/）。

---

背景：Agents Week 全景

@cloudflare/computer 是 Cloudflare Agents Week（8月2-4日）集中发布的一部分，同期还有：

• Agent Development Lifecycle：从 PR 审查到部署的 Agent 驱动 CI/CD
• Cloudflare Agents：统一管理所有部署中的 Agent session 和性能数据
• Cloudflare Wallets：Agent 原生支付和身份（x402 协议）
• Local tracing：wrangler dev 结构化 trace，Agent 可以直接 debug Worker
• Dynamic Workers + Object Capabilities：Worker 间零序列化传递对象引用

整个 Agents Week 的主线是：云基础设施必须为 Agent 原生设计，而不只是把为人类浏览器设计的 Web 基础设施接上 AI。@cloudflare/computer 是这个主线上计算层的核心。

仓库：github.com/cloudflare/computer

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

@cloudflare/computer: Give Every AI Agent Its Own Computer

by Mycelium Protocol

Spinning up a container for every agent has been the default for the past year. Cloudflare’s Agents Week (August 2-4, 2026) named the problem with that approach: there isn’t enough compute on every hyperscaler combined to give 100 million concurrent agents their own containers. Their answer is @cloudflare/computer — an early-preview, open-source agent runtime that dynamically routes execution between cheap Isolates and full Linux Containers, both sharing the same virtual filesystem.

The Core Insight

Most agent work is light: read/write files, check git state, parse JSON, process data. None of that needs a full Linux userland. Cloudflare’s goal with @cloudflare/computer is to keep container invocations below 10% of total execution — with Isolates (Cloudflare Workers) handling the rest.

The architecture that made this possible: agent logic runs in a Durable Object (an Isolate with persistent state and infinite horizontal scale); containers mount on-demand as tools via FUSE. @cloudflare/computer packages this pattern into a single installable abstraction.

Architecture: Shared Filesystem, Dual Backends

The central piece is Workspace — a virtual filesystem backed by SQLite:

Workspace (SQLite-backed virtual filesystem)
   ├── Sources: git repos, cloud storage buckets, arbitrary files
   ├── Isolate backend (fast, cheap)
   │   └── just-bash → JS → Dynamic Worker
   │       Filesystem visible via Worker Binding
   └── Container backend (full Linux)
       └── Cloudflare Containers + FUSE mount
           Changes sync back to Workspace automatically

Both backends share the same files and the same exec(string, options) interface. The agent’s system prompt describes each backend’s capabilities; frontier models are reliable at picking the right one.

Minimal Setup

npm install @cloudflare/computer

Wire a Workspace to a @cloudflare/think agent, expose file/git/shell tools, set the system prompt — the agent can now read, write, run shell commands in isolates, or escalate to a container when it needs npm or native binaries. The Workspace class also provides a node:fs-compatible wrapper for third-party JavaScript libraries.

Why Isolates, Not Just Containers

Cloudflare has been making this bet for almost 10 years:

Year	Milestone
2017	Cloudflare Workers (Isolate model)
2020	Durable Objects (stateful, hibernate, infinite horizontal scale)
2024	SQLite inside Durable Objects (vertical capability added)
2025	Durable Objects can spin up container sandboxes on-demand
2026	@cloudflare/computer wraps all of this into one API

Isolates: millisecond cold start, auto-hibernate when idle (no billing), infinite horizontal scale.
Containers: full Linux, any binary, package managers, vertical scale.
@cloudflare/computer: both, in the same workflow, without forcing a choice.

What Cloudflare Is Already Building With It

• JavaScript apps built, tested, and deployed entirely in Isolates
• Per-customer documentation generated at scale
• Web browser automation for complex tasks

All from the same @cloudflare/computer primitives.

Agents Week Context

@cloudflare/computer shipped alongside Agent Development Lifecycle (AI-driven CI/CD), Cloudflare Agents (unified agent session management), Cloudflare Wallets (x402 agent payments), local tracing for Workers, and Dynamic Workers with object capabilities. The through-line: cloud infrastructure needs to be designed agent-native, not retrofitted from human-facing web infrastructure.

Early preview: github.com/cloudflare/computer · includes step-by-step tutorial at examples/tutorial/

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.