部落格
Two founders, two agents, one plan

How two co-founders in different cities, one using Claude and the other using Codex, can improve the same plan in turn without sending files or merging versions.
Two co-founders can each use the AI agent they prefer and still work on one plan, as long as the plan lives in one shared folder that both agents read and write. In Fabric, that folder is the same for both of them, edits based on an outdated copy are refused and reapplied, and two short Markdown files carry the context from one session to the next. Here is how that plays out over a week.
The founders below are an illustrative example, a composite of how small teams tend to work, and not a Fabric customer.
Two cities, five hours apart
Ana is in Lisbon. She runs product and sales, and she does most of her thinking in Claude, which she reaches through Fabric's connector in the Claude desktop app. Dev is in Toronto. He writes the code, lives in a terminal, and uses Codex for almost everything, including the parts of the job that aren't code.
They're three months into a small B2B tool and preparing a pricing change and a launch for a new tier. The plan for that launch is a single file, launch-plan.md, in a Fabric folder called launch. Next to it sit decisions.md, NEXT.md, a CSV of customer interviews and an AGENTS.md file that tells any agent opening the folder what the project is and how to behave.
Before they set this up, the plan lived in a Google Doc and in each founder's chat history. Ana would ask Claude to rework the pricing section, paste the result into the doc, and message Dev. Dev would copy the doc into a file so Codex could read it, have Codex adjust the rollout steps, and paste them back. Each handoff cost ten minutes and a little accuracy. Twice, a section Ana had rewritten in the morning came back in Dev's evening paste as the old version.
Monday in Lisbon
Ana starts at nine. She opens Claude and asks it to read NEXT.md and launch-plan.md in the launch folder, then rewrite the pricing section using what the interviews say about annual billing.
Claude searches the interview CSV for "annual", finds eleven rows that mention it, reads them, and drafts a new section with two price points and a note about which customers asked for each. Ana edits a sentence, asks for one table to be shorter, and tells Claude to save it. Then she asks it to add a line to decisions.md:
Before she closes her laptop she has Claude update NEXT.md. It replaces the old note with a short one: pricing section rewritten, rollout section untouched, open question about whether existing customers keep their current price.
Nobody sends Dev anything.
Monday afternoon in Toronto
Dev starts at nine his time, which is two in the afternoon in Lisbon. The Fabric desktop app has the shared drive mounted on his laptop, so the launch folder sits in his file system like any other folder, and Codex works in it the way it works in his code repository.
He asks Codex to read NEXT.md first. Codex reads the note, then decisions.md, then the plan, and comes back with a summary: annual billing is in, two tiers, and the rollout steps still describe three tiers. Dev asks it to rewrite the rollout section to match and to add the migration step for existing customers as an open item, since Ana flagged it.
Codex makes the change and saves the file. Dev reads the change, agrees with it, and asks Codex to add his own entry to decisions.md about shipping the billing change behind a feature flag. He ends his day by asking Codex to refresh NEXT.md.
When Ana opens Claude on Tuesday morning, the plan has a new rollout section, a new decision with Dev's reasoning next to it, and a note telling her what's left. She didn't need to ask him what changed.
Wednesday, when both agents edit at once
Most days the time zones keep them apart. On Wednesday they overlap. Ana is working late and asks Claude to tighten the launch email in the plan. At the same moment Dev, early, asks Codex to add a section on support coverage for launch week.
Claude saves first. A second later, Codex tries to save its version of the file, which was based on the copy it read before Claude's edit landed. Fabric refuses that write because the file has changed since Codex read it. Codex re-reads launch-plan.md, sees the new email copy, and adds its support section to the current version. Both changes are in the file, and there's no launch-plan (1).md sitting in the folder afterwards.
This is the part that makes the arrangement safe to leave running. Many sync tools designed for people handle two near-simultaneous saves by keeping the later one, or by keeping both as separate copies. Either outcome is manageable when two people rarely save at once. With two agents writing whenever they're asked to, it would mean lost paragraphs or a folder full of near-duplicates. When two AI agents edit the same file goes through the mechanism in more detail.
Friday, a bad edit
On Friday Ana asks Claude to "clean up the plan before the investor update." Claude is thorough. It removes the open questions section, which Dev was still using to track the customer migration.
Dev notices within an hour. He opens the file's version history in Fabric, finds the version from that morning, and copies the section back. The whole episode takes five minutes, and it leads to a new line in AGENTS.md asking agents to leave the open questions section alone. Dev makes that change himself, since agents can't edit instruction files such as AGENTS.md and CLAUDE.md in Fabric, even ones with write access to the folder. The file every agent reads before it starts is the one file a confused or compromised agent shouldn't be able to rewrite. Your agent instruction files are a security risk explains why.
What made it work
Three things carried the week. Two come from the drive. The third is a pair of habits.
One folder that both agents can reach. Ana's Claude reaches the launch folder through Fabric's MCP server, which also works with ChatGPT, Codex, Claude Code, Cursor and VS Code. Dev's Codex reaches it through the mounted desktop drive, where files take no space on his disk and stream when opened. Different routes, same files. How AI agents connect to a shared drive compares the options, including the CLI and the API. Because the agents meet in the folder, neither founder had to switch tools, and neither agent needed to know the other existed.
Writes checked against the latest version. Wednesday's overlap would have cost one of them a section of work on a drive that lets the last save win. Fabric checks which version each edit was based on and refuses an outdated one, so the agent reapplies its change on top. Searching the interview file also ran on Fabric's side, so Claude could grep the CSV without downloading it first, and an agent that only needs the first rows of a large export can read part of the file. Why AI agents give up searching your files covers the search side.
Two habits. NEXT.md gave each session a starting point, so neither agent began by guessing. decisions.md kept the reasons next to the choices, which meant Dev's Codex could explain on Monday why the plan had two tiers instead of three without anyone writing a summary for it. Both files are plain Markdown and a few lines long. They would help on any shared drive, and they're the subject of how to hand off work between AI agents.
A smaller point: they set up access so that Claude and Codex can each reach the launch folder and nothing else in the workspace. Agent Vault, on the Pro and Team plans, lets you choose the folder, the actions an agent can take, and how long that access lasts. It also logs every read and edit, which is how Dev confirmed on Friday which agent had removed the section.
What it cost them
Very little. Setting up the connector in Claude took Ana a few minutes, and Dev ran one command to add Fabric to Codex and installed the desktop app. Fabric prices per person, so their two agents didn't add seats to the bill; the team plans page has the details, and should AI agents need a seat? explains the reasoning.
The habit change took longer. For the first few days Ana still messaged Dev to say she'd updated the plan. By the second week she stopped, because the note in NEXT.md said it better than her message did.
If you're working with a co-founder this way, the wider argument is in your AI agents need a shared drive, not a group chat, and the basics are in what is a shared drive for AI agents?
Frequently asked questions
Can two people use different AI agents on the same project?
Yes, if the project files live in a folder both agents can reach. In Fabric, one person can use Claude through the MCP connector while the other uses Codex on a mounted drive, and both work on the same files.
Can Claude and ChatGPT or Codex edit the same file?
They can in Fabric. Each agent reads the file, makes its change and saves it, and Fabric checks that the change was based on the latest version before accepting it.
What happens if both agents save the same file at the same time?
Fabric accepts the first save and refuses the second, because it was based on an older version. The second agent re-reads the file and reapplies its change, so both edits end up in one file with no duplicate copy.
How do co-founders share a plan with their AI agents?
Keep the plan as a Markdown file in a shared Fabric folder, connect each person's agent to that folder, and add a NEXT.md handoff note and a decisions.md log beside it.
What should go in NEXT.md?
A few lines on what was finished, what's half done and what you were unsure about. The next person's agent reads it first and starts from there.
What if an agent deletes part of the plan by mistake?
Open the file's version history in Fabric and restore the section from an earlier version. Deleted files go to a trash you can restore from.
Can an agent change AGENTS.md or CLAUDE.md?
No. Fabric protects instruction files such as AGENTS.md, CLAUDE.md and SKILL.md from changes by agents, even when they have write access. People edit them.
Can I limit an agent to one project folder?
Yes. Agent Vault lets you give an agent access to a specific folder, space or project, choose its actions and set how long access lasts. It's on the Pro and Team plans.
Related pages
其他部落格文章:

Agent memory vs a shared drive

Can AI agents use Google Drive or Dropbox?

What happens when an AI agent deletes the wrong file

Why AI agents give up searching your files

Why AI agent access should expire

Your agent instruction files are a security risk

When two AI agents edit the same file

Ask your files, don't build a dashboard