返回列表
🧠 阿头学 · 🪞 Uota学

公司就是文件系统——Agent 的真正基础设施不是 API，是文件

OpenClaw 好用的核心原因是它的上下文就是文件系统；如果把一整家公司建模为文件系统，Agent 就能通过简单的读写文件解决业务问题——文件系统是状态，Claude 是编排器。

2026-02-11 原文链接 ↗
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • 文件系统即状态。 OpenClaw 的架构极简：文件系统存状态，对话本身也是文件，连接的数据（邮件、睡眠）都变成文件。越多数据进入文件系统，Agent 越强大。
  • 企业数据孤岛是 Agent 落地的最大障碍。 发票在 QuickBooks，邮件在 Outlook，合同在 NetSuite——没有共享命名空间。把公司建模成文件系统，Agent 就有了统一的数据访问层。
  • 文件权限天然映射组织架构。 一年级律师助理只能读写自己的案件，合伙人访问所有人的。Unix 文件权限 = 治理结构。简洁到优雅。
  • AI Agent 的本质可以还原为两个组件：文件系统 + LLM 编排器。 剥掉所有噪音，这就是最小架构。

跟我们的关联

  • Uota 的架构就是这篇文章的实证。 ATou 的 OpenClaw 以 ~/clawd 为工作目录，SOUL.md / USER.md / MEMORY.md / memory/ 全是文件。Uota 的能力边界 = 文件系统的丰富度。这不是理论，是 ATou 每天在用的东西。
  • Neta 内部知识管理的启示。 如果把 Neta 的运营流程、品牌规范、数据看板都文件化，Uota（或任何 Agent）就能直接操作业务。当前散落在飞书文档、Notion、各种 SaaS 里的信息孤岛，正是 Agent 发挥不了全力的原因。
  • 「目录即大脑分区」直接呼应 ATou 的 Obsidian 笔记体系。 结构化的文件系统不只是存储，是思维的外化。ATou 作为 Context Engineer，核心工作之一就是设计这个文件系统。

讨论引子

  • Neta 的业务数据散落在多少个系统里？如果要让 Agent 真正能操作业务，第一步该把哪些数据「文件化」？
  • 文件权限映射组织架构听起来优雅，但现实中权限管理远比 Unix rwx 复杂。这个类比的边界在哪？

OpenClaw 之所以好用，核心原因之一是它的整个上下文就是你电脑上的文件系统。

OpenClaw 运行在一台电脑上，让你通过 Telegram 或 iMessage 之类的聊天应用跟它对话。当你让它执行一个任务，它调用 Claude API，使用你机器上的文件作为上下文。你和 OpenClaw 的对话本身也是电脑上的一个文件。当你运行任务，OpenClaw 就往那个文件写入。文件系统就是状态。

随着你往这些文件里添加更多数据，OpenClaw 变得更强大、更好用。当你连接了 Gmail，OpenClaw 就把邮件变成你电脑上的文件。当你连接了 Eight Sleep 智能床垫，OpenClaw 把你的睡眠数据加到文件里。OpenClaw 想接管你的整个世界，但它只有在你的数据进入文件系统之后才能做到。

但如果 OpenClaw 在个人生活中已经这么有用，那如果一整家公司都被表示为它可以操作的文件系统，OpenClaw（或其他 AI Agent）会有多强大？

拿一家律师事务所举例。

在这个世界里，律所被简化为一台电脑上的一组文件夹。

新案子进来，写入 /cases。案子分配给律师，添加到他们的 /cases 文件夹。记录工时，添加条目到 /billing/time-sheet。整个后台运营就是一个状态机。

文件系统的另一个有趣之处是：权限天然映射到组织架构的层级。比如一年级律师助理只对自己的案件有读写权限，而合伙人可以访问所有人的案件。治理结构就是 Unix 文件权限。

企业推行 Agent 之所以复杂，一个原因是数据被孤岛化在许多不同的系统里。发票在 QuickBooks，邮件在 Outlook，提案在 SharePoint，合同在 NetSuite，诸如此类。没有一个共享的命名空间来跨业务访问所有数据。把公司建模成文件系统，Agent 就能访问几乎所有做出正确上下文判断和决策所需的数据。

当然，所有企业都有各种微妙之处，很多工作流程编码在人的脑子里——不在 JSON 文件里。但 OpenClaw 及其底层架构指向的未来是：文件系统成为对 Agent 最有用的那个真相之源。

过去一年 AI Agent 爆发式增长。但当你剥掉噪音，AI Agent 的架构可以被还原为两个组件：文件系统作为状态，Claude 作为编排器。把公司建模成文件系统，Agent 就能通过简单的读写文件来解决业务问题。

  • 原文：https://x.com/mernit/status/2021324284875153544
  • 作者：mernit
  • 发布：2026-02-10
  • 翻译：Opus

One of the reasons Openclaw is so good is because its entire context is a filesystem on your computer.

Openclaw runs on a computer and lets you talk to it via a chat app like Telegram or iMessage. When you ask it to run a task, it calls the Claude API and uses context from files on your machine. Your conversation with Openclaw is represented as a file on the computer. When you run a task, Openclaw writes to that file. The filesystem is the state.

As you add more data to those files, Openclaw becomes more powerful and useful. When you connect your Gmail, Openclaw has emails as files on your computer. When you connect your Eight Sleep bed, Openclaw adds your sleep data to a file on the computer. Openclaw wants to take over your world, but it can only do that if your data is in the filesystem.

OpenClaw 之所以好用，核心原因之一是它的整个上下文就是你电脑上的文件系统。

But if Openclaw is useful for our personal lives, how powerful would Openclaw (or other AI agents) actually be if an entire company was represented as a filesystem it could work in?

OpenClaw 运行在一台电脑上，让你通过 Telegram 或 iMessage 之类的聊天应用跟它对话。当你让它执行一个任务，它调用 Claude API，使用你机器上的文件作为上下文。你和 OpenClaw 的对话本身也是电脑上的一个文件。当你运行任务，OpenClaw 就往那个文件写入。文件系统就是状态。

Let's take a law firm, as an example.

随着你往这些文件里添加更多数据，OpenClaw 变得更强大、更好用。当你连接了 Gmail，OpenClaw 就把邮件变成你电脑上的文件。当你连接了 Eight Sleep 智能床垫，OpenClaw 把你的睡眠数据加到文件里。OpenClaw 想接管你的整个世界，但它只有在你的数据进入文件系统之后才能做到。

In this world, the law firm is reduced to a set of folders on a computer.

但如果 OpenClaw 在个人生活中已经这么有用，那如果一整家公司都被表示为它可以操作的文件系统，OpenClaw（或其他 AI Agent）会有多强大？

When a new case comes in, we write to /cases. When the case is assigned to a lawyer, we add the case to their /cases folder. When they log time, we add the entry to /billing/time-sheet. The entire back office operation is just a state machine.

拿一家律师事务所举例。

Another interesting aspect of the filesystem is that permissions naturally map to seniority in the org chart. For example, a first-year associate gets read/write access on their cases, whereas partners can access everyone's cases. The governance structure is just unix file permissions.

One reason that rolling out agents at enterprises is complicated is because data is siloed across many different systems. Invoices are in Quickbooks, emails are in Outlook, proposals live in Sharepoint, contracts live in Netsuite, and so on. There is no shared namespace to access all this data across the business. By modeling a company like a filesystem, agents can access nearly all the data they need to get the right context and make decisions.

在这个世界里，律所被简化为一台电脑上的一组文件夹。

There's obviously nuance to all businesses, and many work streams are codified in people's heads - not in JSON files. But the power of Openclaw and the underlying architecture points to a future where the filesystem becomes the source of truth for the agents that are the most useful.

新案子进来，写入 /cases。案子分配给律师，添加到他们的 /cases 文件夹。记录工时，添加条目到 /billing/time-sheet。整个后台运营就是一个状态机。

The past year has been explosive for AI agents. But when you tear away the noise, the architecture of an AI agent can be reduced to two components: the filesystem as state, and Claude as the orchestrator. By modeling the company as a filesystem, an agent is able to solve business problems by simply reading and writing files.

文件系统的另一个有趣之处是：权限天然映射到组织架构的层级。比如一年级律师助理只对自己的案件有读写权限，而合伙人可以访问所有人的案件。治理结构就是 Unix 文件权限。

企业推行 Agent 之所以复杂，一个原因是数据被孤岛化在许多不同的系统里。发票在 QuickBooks，邮件在 Outlook，提案在 SharePoint，合同在 NetSuite，诸如此类。没有一个共享的命名空间来跨业务访问所有数据。把公司建模成文件系统，Agent 就能访问几乎所有做出正确上下文判断和决策所需的数据。

Link: http://x.com/i/article/2021308996020211712

当然，所有企业都有各种微妙之处，很多工作流程编码在人的脑子里——不在 JSON 文件里。但 OpenClaw 及其底层架构指向的未来是：文件系统成为对 Agent 最有用的那个真相之源。

过去一年 AI Agent 爆发式增长。但当你剥掉噪音，AI Agent 的架构可以被还原为两个组件：文件系统作为状态，Claude 作为编排器。把公司建模成文件系统，Agent 就能通过简单的读写文件来解决业务问题。

Your Company is a Filesystem

  • Source: https://x.com/mernit/status/2021324284875153544?s=46
  • Mirror: https://x.com/mernit/status/2021324284875153544?s=46
  • Published: 2026-02-10T20:43:58+00:00
  • Saved: 2026-02-11

Content

One of the reasons Openclaw is so good is because its entire context is a filesystem on your computer.

Openclaw runs on a computer and lets you talk to it via a chat app like Telegram or iMessage. When you ask it to run a task, it calls the Claude API and uses context from files on your machine. Your conversation with Openclaw is represented as a file on the computer. When you run a task, Openclaw writes to that file. The filesystem is the state.

As you add more data to those files, Openclaw becomes more powerful and useful. When you connect your Gmail, Openclaw has emails as files on your computer. When you connect your Eight Sleep bed, Openclaw adds your sleep data to a file on the computer. Openclaw wants to take over your world, but it can only do that if your data is in the filesystem.

But if Openclaw is useful for our personal lives, how powerful would Openclaw (or other AI agents) actually be if an entire company was represented as a filesystem it could work in?

Let's take a law firm, as an example.

In this world, the law firm is reduced to a set of folders on a computer.

When a new case comes in, we write to /cases. When the case is assigned to a lawyer, we add the case to their /cases folder. When they log time, we add the entry to /billing/time-sheet. The entire back office operation is just a state machine.

Another interesting aspect of the filesystem is that permissions naturally map to seniority in the org chart. For example, a first-year associate gets read/write access on their cases, whereas partners can access everyone's cases. The governance structure is just unix file permissions.

One reason that rolling out agents at enterprises is complicated is because data is siloed across many different systems. Invoices are in Quickbooks, emails are in Outlook, proposals live in Sharepoint, contracts live in Netsuite, and so on. There is no shared namespace to access all this data across the business. By modeling a company like a filesystem, agents can access nearly all the data they need to get the right context and make decisions.

There's obviously nuance to all businesses, and many work streams are codified in people's heads - not in JSON files. But the power of Openclaw and the underlying architecture points to a future where the filesystem becomes the source of truth for the agents that are the most useful.

The past year has been explosive for AI agents. But when you tear away the noise, the architecture of an AI agent can be reduced to two components: the filesystem as state, and Claude as the orchestrator. By modeling the company as a filesystem, an agent is able to solve business problems by simply reading and writing files.

Link: http://x.com/i/article/2021308996020211712

📋 讨论归档

讨论进行中…