The folder everyone has
Open the shared folder of almost any team and you will find it within a minute. Somewhere near the top, sitting among the pitch decks and the receipts, is a plan.
Not one plan. A small family of them. plan.md, plan (1).md, plan copy.md, plan-final.md, plan-final-v2.md, and at the bottom, in capitals, like someone raising their voice in a meeting, plan-final-v2-ACTUAL.md.
I have run a few businesses, and I have made this folder in every one of them. I used to think it was a filing problem, something a tidier person would not have. I now think the names are a kind of fossil record, and that each one tells you exactly what happened on the day it was made.
Reading the layers
plan.md is the founding document. Someone wrote it on a Monday with good intentions and a clear head. Nothing is wrong with it, except that it is the oldest thing in the folder and therefore, by some law of offices, the first thing people open.
plan (1).md is an accident. It means somebody downloaded the file when a copy already sat in their Downloads folder, and the operating system politely made room. Nobody chose that name. It was simply what a collision looks like, once it has been made polite.
plan-final.md was final for about four hours. It was named in a burst of optimism by a person who had just finished a long edit and wanted the matter closed. Then someone read it and had a thought. Hence plan-final-v2.md, which has the comments from the call, and which was also final.
And then plan-final-v2-ACTUAL.md. This is the interesting one, because the word in capitals is not for you. It is a message to the person who opens the wrong file next week. It says: I have been burned by this folder, and I am leaving a note. It is the only honest file name in the building, and it quietly admits that the system has failed.
Why a copy is the default
None of this is the fault of the people. It comes from how most cloud storage works. You download a file, or a sync client does it for you, you change your copy, and you upload it back. The whole arrangement rests on a quiet hope that nobody else did the same thing in between.
When somebody did, the system has two bad options. It can overwrite one person's work with the other's, which is quick and loses things. Or it can keep both and give one of them a new name, which is polite and gives you plan (1).md. Most products choose the second, because losing work gets you angry emails and clutter does not.
So every collision leaves a file behind. A team of six, over a year, leaves a great many. And the files are not even wrong. Each one was correct on one laptop at one moment. The trouble is that a file which is correct on a laptop is not the same thing as a file the team is actually using, and there is nothing in the folder to tell you which is which.
The same bet gets placed with no person involved at all. A script that downloads a file, transforms it and uploads the result carries the same quiet hope, only it places the bet far more often than a human would, and it never notices when it loses. Nobody is even there to feel the pang.
Then we gave everyone an agent
For a while this was a manageable mess, because people are slow. Making a stray copy takes a person a few clicks, and they feel a small pang about it, which keeps the numbers down.
Agents feel no pang. They are fast, they are tireless, and they are very good at doing what you asked. Ask one to tighten the goals in the plan and it will open the file, do a careful job and save the result, in the time it takes you to refill your coffee. Ask three of them to do three different things to the plan, in three people's working copies, and you will have more versions by lunch than the old method produced in a quarter.
What worries me is the diligence. Each agent produces a clean, well-reasoned version of the plan based on the plan as it looked when it started. None of them knows the other two exist. All three are right, and the folder now holds three truths with no way to say which one wins.
I have watched a person spend a morning reading three agent-edited plans side by side, working out which sentences came from which. An agent should not be creating that kind of work for a human. Saving people work was the whole idea.
What we tried first
We did the sensible things, because we are sensible people. We agreed a naming convention. We agreed that only one person edits the plan, and the others send suggestions. We agreed to put the date at the front, which lasted eleven days.
Every one of these asks a person to do the job of a system, and the rules do not survive a busy week. You can write the convention in a file and the agent will, to be fair, usually follow it. It still does not help, because the naming was never what went wrong. Two people could each hold a private copy of the same thing, and neither of them could see theirs going stale.
The date at the front was the cruellest one. It told you when a file was made and nothing about whether it was still the one to trust. A file can be three weeks old and perfectly current, or three minutes old and already out of date. The name has no way to know, because the name is not connected to anything.
One file, and the rule that guards it
The fix is smaller than it sounds. There is one plan.md. Not a master copy and some working copies: one file, in one folder, which everybody opens, whether that somebody is a person, an agent, or an agent being run by a person in another city. Mounted as a folder, over MCP, from the command line, it is the same file in each case.
Every save becomes a new revision of that file, so the history is kept and the name stays put. And every edit has to say which revision it was based on. If the file has moved on since the agent read it, the save is refused with stale_version, and nothing is written at all.
What happens next is the good part. The agent does not panic and does not make a copy called plan-agent-retry.md. It reads the latest version, looks at what changed, and redoes its own change on top. Two edits that would once have fought now land one after the other, in order, and neither person was asked to referee.
It is a dull rule, which is why I trust it. A sensible colleague does the same thing by instinct before saving over somebody else's work: hang on, this has changed since I looked.
And when it all goes wrong anyway
Some of the clutter in these folders is not collisions at all. It is fear. People keep plan-old.md and plan-before-the-call.md because they do not trust that a deleted file can come back, and they are right not to, in most places.
So we made deletion forgiving. A deleted file goes to the drive's trash, and for an hour you can put it back. That is not a long time, and it is not meant to be an archive. It is just enough time to notice that an agent tidied up a little too enthusiastically, which, if you have ever watched one work, you will know happens.
I would like to tell you that this means we no longer name things badly. We do. Last week I created a file called notes-2.md, and I am not proud of it. But the plan is one file now, and when I open it I am not left to wonder whether I am looking at the right one.
I still think about that ACTUAL in capitals, though. Somebody typed it with a small sigh, hoping to save a stranger the trouble. It was a kind thing to do, and nobody should have needed it.