Multi-agent, with one person
A lot of what gets called multi-agent work is solo work with extra steps. One person, a few agents, one laptop. The researcher agent writes notes, the coding agent reads them, a third agent checks the result, and it all hangs together beautifully.
I do not think it hangs together because of clever orchestration. It hangs together because everything those agents touch is in one place, which is the laptop. They share a disk. They can see each other's output the way people in the same small room can hear each other think.
Nobody ever had to design that. It came free with the setup, so we tend not to notice it is doing the work.
Then somebody joins
Now picture the next thing that happens, which is that a second person wants in. They are good at their job and they have opinions about their tools. They use a different agent from yours, because everyone has a favourite and nobody is going to change it for you. And they live somewhere else, because of course they do.
Suddenly the free thing is gone. Your agents' notes are on your disk and theirs are on theirs. The usual answers arrive in order: send the file over, paste the summary into a chat, put it in a shared drive and hope.
I have been through every one of those, and I can tell you where they end up. They end up with a person standing between two agents, carrying messages. Which is a strange job to give someone who bought the agents to save time.
Some teams respond by making the agents formally talk to each other, and I understand the appeal. But it adds a second thing to maintain, and it still leaves the question of where the actual plan sits while they are talking.
The afternoon it clicked
I am the CEO of Evokoa, and my co-founder Dalton Prescott is the CTO. Dalton lives in one city and I live in another. He works with Claude Code and I work with Codex, and neither of us has ever been tempted to swap.
We had a launch plan, one Markdown file in a shared folder, with goals at the top and a long list of tasks underneath. One afternoon Dalton asked Claude Code to tighten the goals. They had got baggy, the way goals do when five people have added to them, and he wanted them sharper and fewer.
A little later, with no idea he had done that, I asked Codex to give every task in the plan an owner. I wanted to see at a glance who was holding what.
Codex opened the plan, and the goals were already different. It did not ask me about it. It kept the tightened version, added the owners, and saved. I read the result and could not tell where one of us ended and the other began, which was the point.
What did not happen
Consider what nobody did. Nobody sent a file or pasted anything into a chat. Nobody opened a merge screen to pick between two versions of the goals. Nobody said hold on, let me finish before you start.
And the two agents never spoke. Claude Code did not know Codex existed. Codex had no idea that anyone had changed the goals that afternoon, only that the file it opened looked the way it did. Each did the job it was asked and stopped.
I keep coming back to how little coordination there was. It is tempting to think agents need protocols and messages and handoffs to cooperate. Maybe some jobs do. This one needed what every office already has: both parties looked at the same page.
Shared context is a place
When people say they want shared context for their agents, they often describe it as a feature. A memory layer, a sync setting, an integration. I have started to think that framing is the wrong way round.
Context is what is lying around when you sit down to work. A desk with last night's notes on it has context. A kitchen table with the plan pinned to the wall has context. You do not call it a feature of the table. It is just where you are, and everyone who sits there inherits it.
Agents are the same. They are trained on a great deal of reading, searching and editing files, so a folder is already the interface they know best. If the folder is shared, the context is shared. If it is not, no amount of clever messaging in between will quite make up for it.
That is the thinking behind Baseseat, the multiplayer folder we make: one folder that a whole team and all its agents work in at the same time, reachable through Finder, a terminal, MCP, the CLI or the Python SDK. I would rather not oversell it. It is a folder. That was the idea.
Why two edits did not collide
You might fairly ask what would have happened if both agents had saved at the same moment. The honest answer is that one of them would have been told to try again.
Every save of a file makes a new revision, and an edit has to say which revision it was based on. If the file changed in the meantime, the write is refused with stale_version and nothing is written. The agent then reads the latest version and redoes its change on top of it. Nobody overwrites anybody, and both changes land in order.
In our case the timing was kind and the second agent simply found the first one's work waiting. But I like knowing that the unkind version has an answer too.
If your team is starting out
You do not need our product to try the idea. You can begin this week, with a plain folder and a few habits. Keep plans and decisions as plain Markdown, in one folder everybody can reach. Agents read it well, people can read it without a special app, and it does not belong to anybody's tooling. A decisions.md that says what you chose and why, in a few lines each, is worth more to the next agent than a long chat transcript.
Let each person bring the agent they like. This is easier than it sounds once the work lives in a shared place, because the agent stops mattering as much. What matters is that all of them read the same files.
And write things down where the next agent will look. That is the habit that pays back fastest. Agents do not remember yesterday, but they will read a file called NEXT.md at the top of the folder whenever you tell them to, and you can tell them once. If you finish a day by leaving a few honest lines about what is half done and what you were worried about, the next agent, yours or someone else's, starts the day knowing it. I was sceptical of the last one at first, since it sounds like tidiness advice. It stopped sounding like that the first time an agent picked up a half-finished job from a note I had left in a hurry and finished it correctly.
One more thing, because it is easy to forget: read what the agents did. A shared place makes their work visible, and visible work is only useful if somebody looks. We skim the plan each morning, the way you would glance at a whiteboard.
On cost, since someone will ask: it is $20 a person and $0 an agent, so adding another agent never costs you a seat.
Back to the desk
I still think about how ordinary that afternoon was. Dalton and I did not discuss it. We found out afterwards, reading the file, and had a short laugh about it on a call.
We have been working with people in other cities for years, and what has always made that work is a shared place: the same board, the same document, the same thread. Agents turned out to want exactly the same thing.
They were already working as a team. They just needed a desk.