Not subscribed? Sign up to get it in your inbox every week.

Three years into the AI boom, no one has really solved collaboration. This summer, everyone started trying at once.

Chat was you and a window. Cowork was you and an agent in a private session, working your files on your machine. The tools improved by orders of magnitude, but the shape never moved: one person, one AI, a closed door between that session and the rest of the company. Whatever the model learned about your work died in your chat history, while the person one seat over taught their own agent the same things from scratch.

You can already watch a version of this run. On the website of a startup called Axamy there's a page of fictional companies you can boot: a media conglomerate called Waystar Royco, a biotech called Lumon Industries (both will ring a bell if you've watched HBO lately), a resort in Koh Samui named White Lotus, a maple syrup operation called True North Maple, and, for the ambitious, the Department of Defense. Pick one and it boots into a live demo: an AI-run workspace where goals turn into plans, work gets assigned to the right fictional people, check-ins run on their own, and follow-ups get chased without anyone doing the chasing.

This company caught my attention because of the thesis: coordination itself is the product. You drop a goal wherever your team already talks, the goal becomes a tracked plan with owners and follow-ups, and the agent checks in with each person separately, surfacing only the decisions that need human judgment.

One line from the founder's essay on becoming post-bureaucratic stuck with me: "your chats are yours, but everyone on the team is already talking to the same Axamy."

That sentence is the whole category. Every private conversation feeds one shared system, and that system holds the goals, the status, the process, and the preferences in a single place. The process gets encoded once instead of living in nine heads and four personal chat histories, which is the standardization ops leaders have been chasing with wikis and SOPs for decades, except this time the document does its own enforcement. The shuttling between people, the assigning and reminding and reconciling that fills a middle manager's day, becomes the machine's job. That job has been growing, too: by the essay's count, the middle-manager share of the workforce has roughly doubled since the 1980s.

The operator math is simple and slightly cruel. A single-player agent multiplies one person's output, and coordination load grows with that output: more drafts to review, more decisions queued, more threads to reconcile with each other. One line from the same startup's blog puts it plainly: keep running the management processes that worked when you were doing 10 things, and you start dropping things at 40. The faster the individuals get, the more the seams between them become the constraint.

Stanislas Polu, Dust's cofounder, sees the same thing from the enterprise side. As agents take on tasks that run for days or weeks, "you no longer collaborate with an agent by yourself." His sharper line is one I would happily frame on the wall of this newsletter: "The bottleneck shifts from generation to coordination." In May I argued that cheap execution moved the bottleneck to strategy. Coordination is the other half of that same shift: once everyone's output triples, pointing it all in one direction becomes the scarce work, and the single-player tools have nothing to say about it.

There's also an ownership angle, which finishes an argument from July. When I wrote about owning your intelligence stack, the claim was that models are rented while context, skills, and playbooks can be owned. A multiplayer workspace is where the owned layer gets an address. A company whose AI context lives in personal chat histories owns none of it, and the day someone leaves, their agent's education walks out the door with them. The shared workspace turns that education into a company asset by default, the way Ramp built by hand and Claude Tag productized for Slack. This new wave is trying to productize it for the whole org chart.

Pick one project with at least three people on it, and ask each of them to show you their AI conversations about it. Count the number of times the same context got explained from scratch to an agent a teammate had already taught. That count is your coordination tax, and it's the exact number these products exist to drive to zero.

Then run the cheap experiment. Move one shared goal into a surface every teammate's agent can see (a Tag channel, a shared project, a pilot of one of these workspaces) and watch whether the re-explaining stops. Keep the June discipline while you do it, because multiplayer raises the stakes in both directions: a shared agent that drifts is wrong in front of everyone, and an agent that runs your coordination will faithfully standardize whatever process you hand it, including a bad one. Centralize on purpose, or you will centralize by accident.

For three years, AI met us one at a time. It's starting to meet the team.