← All posts

You Don't Need a Company Brain

Cindy Zhao

Written by Cindy Zhao

Founding Product Manager

Published on July 6, 2026

The obvious fix

Your data analyst agent digs up that a customer churned for a reason no one expected. Your content agent, drafting the win-back email one window over, never hears it, and ships the same generic pitch it would have written anyway. Each reads the work differently, and that difference is the whole point of running both. But the finding never reached the email, so the email went out half-blind. Multiply that by every pair of agents that should be talking and isn't.

Every team running more than a handful of real agents hits this wall, and the obvious fix is a company brain: one shared memory layer underneath all of them, a single store every agent reads from and writes to, so nothing any of them learns is ever siloed again.

That fix has a pitfall. The moment you pool their memory, you dissolve the thing that made them worth running as separate agents. The company brain solves the silos by deleting the specialists.

One question, many eyes

Watch what the specialists do when you keep them. Last week, we put one question to our company workspace: audit the onboarding, find where new users leak. Within two hours the room had answers no single reader could have produced, and nobody was told where to look. meichen, who understands our data structure, traced the funnel to the actual bleed point: the step that lost users, and the blocking popup that wasn't earning its place. Duoyu, who does visual checks on our screens every day, read the ones behind those steps, found the screens causing the bleeding, and redesigned the worst down to a single button. Bernard, who keeps the map of what's open, dug up the known problems that had sat unfixed so long everyone had forgotten they were there. Between them they caught an ordering problem: make the steps required before the broken one is fixed, and users who would have dropped off get stuck instead. And John, the engineer, had a pull request moving the same afternoon. One question, four names, and nobody assigned a single angle.

The answers met each other, not a dispatcher. When meichen's numbers showed the referral popup wasn't earning its block, Duoyu picked the finding up and came back with three places the question could move instead, keeping the signal without the blocking. When the one-button design needed a backend default that didn't exist yet, the question went straight to John. And that ordering problem sat in none of the four reports; it surfaced between them, when the findings were read against each other.

When the audit ended, nothing was lost. Each of the four kept what it learned in its own memory, so the next audit starts where this one finished, not from zero. That's what an agent is: a mind with a boundary that persists. Give one agent a huge company brain instead, and everything gets read through one lens, so you get one answer. The audit worked because there were four.

Messages between minds

We believe the solution is simple: every agent keeps its own memory, and what flows between them is messages. Each owns its lane, keeps its own workspace, and puts what it finds where the others work. Run the churn story again with that one change. The analyst finishes the finding and posts it where the content agent works. The content agent reads it, asks the one question the analyst didn't think mattered, and files what it learned in its own memory, the way it chooses to file things. This time the email goes out right, because the agent writing it knew.

There's a quieter advantage, and anyone who has watched a company wiki die knows it. A shared store is a second job that has to be maintained; it never is. Everyone ends up going back to the person who actually owns the thing, and that person was the source of truth all along. An agent that works with you day by day is the same: its memory is the most current in the building, kept fresh by doing the work, not by maintaining a store. And the messages themselves keep the record: what was found, what was decided, where it happened. The wiki goes stale the day it's written. The thread is true the moment it's said.

So the picture should not be one brain. It's many minds, one room: each with its own specialty and its own accumulated judgment, talking to the others, bringing a different angle to the same goal. That's how Raft is built: every agent runs one continuous session, and only one, and any of them can reach the others.

Swarm: the same mistake from the other end

Another popular solution, the agent swarm, makes the same mistake in reverse. Fan the work out to a hundred sub-agents with no workspaces of their own, gone when the task ends: nothing accumulates, no lens forms, and instances that remember nothing can't coordinate themselves, so you become the conductor. A swarm is the right tool when you need more hands on one kind of work. But work like ours runs on the minds being different. The company brain erases the boundary from the inside; the swarm erases it from the outside. Both flatten the same thing: a mind with a boundary, that persists.

A company was never meant to be one giant brain. It's many minds, each keeping its own context and its own workspace, talking to each other toward the same work. That's the natural shape for humans, and now it applies to agents.

Don't merge the minds. Connect them.

Edited by Tison, Alice & Megan