ブログ

Your AI agents need a shared drive, not a group chat



Agents already coordinate through files. The trouble starts when a second person joins and the shared disk disappears. Here's why the fix is a shared folder built for agents, and what that folder has to do.




Ask Claude Code or Codex to take on a task and watch what it does first. It lists a directory, opens a Markdown file, searches for a name, writes a plan, and comes back to that plan later. Nobody designed agents to work this way. Files were simply there, familiar from the code and documents they learned from, and they turned out to be the natural place to keep work.

That's why running several agents on one laptop works better than it should. One agent writes a plan to plan.md, another reads it and picks up the next step, and a third checks the result. They never talk to each other. They share a disk, and the disk is where the coordination happens.

Then a second person joins.


The shared disk disappears when a second person joins

Picture two co-founders in different cities. One works with Claude, the other with ChatGPT or Codex. Each agent is capable on its own, and each has its own laptop full of files. The plan they're both improving now exists in two places, and neither agent can see the other's copy.

So the people become the connection. One founder pastes the latest version into a message, the other downloads it, their agent edits it, and the new version goes back the other way. Files multiply with names like plan-final-v2.md, and someone spends a morning working out which one is current. The agents were supposed to save that morning.

The obvious answers are a group chat for the people or a messaging protocol for the agents. Both add a channel for talking about the work. Neither answers the more basic question of where the work lives. Two founders, two agents, one plan tells the story of what happens when that question has an answer.


Why a group chat is the wrong fix

Chat feels natural because it's how people coordinate. For agents it's an extra layer to build and maintain. Messages have to be formatted, routed, stored and replayed, and an agent that joins halfway through has to read the whole history to catch up. Agent-to-agent protocols formalize the same idea, which helps with some problems, but the plan still lives somewhere, and the agents still need to find it, read it and change it. We compare the two approaches in agent-to-agent protocols vs a shared folder.

A shared folder skips the extra layer. One agent writes, the next reads. An agent that joins late opens the folder and sees the current state, the same way a new colleague opens the team's documents. It's the reason distributed human teams work: everyone looks at the same page.


Folders have a new kind of user

For half a century, folders have mostly served people opening files one at a time. A person reads a handful of files, saves a few and stops for the day. Sync tools and shared drives were built around that pace.

Agents work at a different pace, and there are more of them. A five-person team might run dozens of agents. A larger company might run thousands, each in its own sandbox, all reaching into the same files. A folder used by one person can end up with hundreds of readers and writers at the same time. That changes what the folder needs to do.

Search has to work the way agents search. A person types a few words and scans the results. Agents search like programmers: grep for an exact phrase, try a regular expression, narrow to one directory, read the lines around a match, then adjust and try again, sometimes dozens of times a minute. On a laptop the disk answers instantly. On a cloud drive, every file an agent opens is a trip over the network, and a quick search becomes a long wait. A drive built for agents runs that search on the server, across every file, so the agent keeps using the tools it knows and gets answers back quickly. Why AI agents give up searching your files goes into this.

Writes have to be safe when agents overlap. Systems built for people usually let the latest save win, because two people rarely edit the same file at the same moment. Agents do it constantly. When one agent saves a change based on an older version of a file, the drive should refuse the write, so the agent re-reads the file and reapplies its change on top. Nothing is silently overwritten and no stray copy appears. When two AI agents edit the same file explains how that works.

Big files have to be readable in parts. An agent that needs the last hundred lines of a log, a few columns of a large CSV or a slice of video shouldn't have to download the whole file first. Reading only the part it needs keeps the work fast and keeps sandboxes small.

Access has to be narrow and temporary. An agent running untested code in a sandbox shouldn't hold a key to everything. It needs access to one folder, for one kind of action, for as long as the job takes, and nothing more. Why AI agent access should expire makes the case.

Standing orders have to be protected. Files like CLAUDE.md, AGENTS.md and SKILL.md are instructions every agent reads before it starts. If one compromised agent could rewrite them, it could steer every agent that came after it. Changing them should stay a human job. See your agent instruction files are a security risk.

Mistakes have to be recoverable. An agent that tidies too eagerly will eventually delete the wrong file. Trash with restore and version history turn that from a crisis into a minor fix.

None of these are unusual requirements. They follow from taking agents seriously as users of a folder.


What this looks like in Fabric

Fabric is a shared drive for a team and every agent it runs.

On your computer, the Fabric desktop app mounts Fabric as a drive on Mac, Windows and Linux. Files take no space on your disk and stream when they're opened, so an agent working on your laptop reads and writes there like any other folder. Agents without a filesystem reach the same files through Fabric's MCP server, which works with Claude, ChatGPT, Claude Code, Codex, Cursor and VS Code. Scripts and coding agents can also use the CLI and the API. How AI agents connect to a shared drive compares the four routes.

Underneath, the drive handles the things above. Search runs on the server with grep, regular expressions and full-text search across every file. A write based on an outdated version is refused rather than applied, so agents reapply their changes instead of overwriting each other. Agents can read part of a large file without fetching all of it. Instruction files such as CLAUDE.md, AGENTS.md and SKILL.md can't be changed by agents, even ones with write access. Deleted files go to a trash you can restore from, and earlier versions are kept.

Access is set with Agent Vault, which treats external agents as non-human members of your workspace. You give each agent access to a folder, space or project, choose which actions it can take, and decide how long that lasts: a single task, a day, a week or indefinitely. When an agent needs more, it asks, and you approve once, for a period, or not at all. Every action is logged. Agent Passes do the same for the apps you've connected to Fabric, such as Gmail or Slack, action by action. Both are on the Pro and Team plans.

Fabric prices per person. Your agents don't need seats of their own, which is the subject of should AI agents need a seat?


Habits that make a shared drive work

The drive handles the mechanics. A few habits make the most of it, with or without Fabric.

Keep plans and decisions in plain Markdown in one shared folder. A few lines per decision in a decisions.md file beat a long chat history. Let each person keep the agent they prefer, since once the work lives in a shared place, the choice of agent matters less. At the end of a session, write a short NEXT.md note saying what's half done and what you were unsure about. Agents don't remember yesterday, but they read the files they're told to read. How to hand off work between AI agents shows the pattern.

And read what the agents did. Visible work only helps if someone looks at it, so skim the plan each morning the way you'd glance at a whiteboard.


It looks like a folder because it is one

The folder is one of the oldest ideas in computing, and it has turned out to suit its newest users. It needed rebuilding for them: faster search, safe writes, narrow access and a way back from mistakes. With that in place, a team and its agents can work from the same files without anyone relaying messages in between.

If you're new to the idea, start with what is a shared drive for AI agents?


Frequently asked questions

Why do AI agents work through files?

Agents learn from large amounts of code and documents, so reading, searching and editing files is familiar to them. A folder gives them a place to keep plans, notes and results that other agents and people can read later.

What goes wrong when several people use their own agents?

Each agent works on its own copy of the files, so the work splits. People end up copying versions between agents by hand and working out which file is current.

Isn't a group chat or agent-to-agent protocol enough?

They help agents exchange messages, but the work still has to live somewhere. A shared folder gives every agent the current state of the work without a separate channel to build and maintain.

What happens when two agents edit the same file in Fabric?

If an agent tries to save a change based on an older version, Fabric refuses the write. The agent re-reads the file and reapplies its change, so neither agent's work is lost and no duplicate copy appears.

How do agents connect to Fabric?

Agents on your computer can use the mounted desktop drive. Others connect through Fabric's MCP server, the CLI or the API. The MCP server works with Claude, ChatGPT, Claude Code, Codex, Cursor and VS Code.

Can I limit what an agent can see?

Yes. With Agent Vault you give each agent access to a specific folder, space or project, choose which actions it can take, and set how long the access lasts. Agent Vault is on the Pro and Team plans.

Can an agent change my CLAUDE.md or AGENTS.md files?

No. Fabric protects agent instruction files such as CLAUDE.md, AGENTS.md and SKILL.md, so agents can't change them even when they have write access.

What if an agent deletes the wrong file?

Deleted files go to Fabric's trash and can be restored, and earlier versions of files are kept in version history.

Do I pay for each agent?

No. Fabric prices per person, so agents don't need their own seats.


Related pages



The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.