The workspace assumption.

When I tell people I left one of the world's largest collaboration companies to build communication software for agents, the usual inference is that I believe agents are about to replace most of the workforce. It is a reasonable inference, and I want to reject it clearly, because it obscures the actual argument. I do not think agents are about to replace teams. I think they have already changed what a team is, and the software teams coordinate through has not noticed.

Nearly every tool a team runs on encodes the same silent premise: every participant is a person. I call this the workspace assumption.

The assumption was so reliably true for fifty years of software that nobody thought of it as an assumption at all. Tickets compensate for what people forget. Standups compensate for what people cannot see. Docs compensate for what no one can hold in their head. My claim is that the premise underneath all of this stopped being true sometime in the last two years, and that much of the friction companies now feel with AI is that fact surfacing.

Let me state the terms of the argument before pressing it. It assumes agents keep improving at multi-step work across tools at roughly the rate we have seen recently; it does not require a breakthrough. It assumes the unit of agent output keeps shifting from answers to finished work. I am confident in the direction of both trends and much less confident in their pace. And there is an honest falsifier: if agents plateau as single-shot answer machines, the workspace assumption holds and this essay overstates its case. I think that outcome is unlikely, but the argument depends on it not happening, and you should know that.

The mechanism runs in three steps. The capability step: agents can now read a codebase, operate software, hold state across a task, and return completed work rather than a suggestion. The structural step: that capability lives almost entirely inside single-player surfaces, a terminal, a personal chat window, an editor, where each session accumulates context the organization cannot see or reuse. The consequence: every agent added to a team adds real output and real coordination work in the same motion, because a person must carry context in and carry results out. The coordination surfaces cannot help; they have no way to represent an agent as a participant, only as a topic people talk about.

A concrete version. An engineer hands a bug investigation to an agent. It reads the relevant code, pieces together how the code got that way, checks whether related incidents exist, and comes back twenty minutes later with an explanation and its sources. This already happens today. Now follow the output. It sits in one person's terminal. The teammate debugging the adjacent service cannot find it. The incident review three weeks later will not cite it. If the engineer is out on Thursday, the organization does not know the investigation ever occurred. The capability in that story is brand new. The organization around it is unchanged.

The strongest objection is that incumbent collaboration products will simply add agents, and it deserves a serious answer. The incumbents have distribution, organizational data, and years of workflow gravity, and they will ship agent features aggressively. But an agent dropped into a channel still has no persistent identity, no scoped authority, no reusable context, no way to be delegated to, no evidence trail, and no way to coordinate with other agents. The interface is the easy tenth of the problem. A second objection says this is premature because agents are unreliable, and it is right about today's failure modes. The useful move is to sort the constraints: reliability and cost improve with model capability; identity, attribution, and shared context do not, because they are properties of the surrounding system rather than the model. The second kind of constraint is the kind you have to build for deliberately.

The descriptive fact is that agent capability is advancing faster than organizational integration. My prediction, held with moderate confidence on timing, is that this gap becomes one of the largest constraints on enterprise AI adoption within a few years. The recommendation does not wait on the prediction: move agent work into shared, attributed, inspectable surfaces now, while the stakes are low, because organizational habits change more slowly than model capabilities do.

The workspace assumption was never really about software. It was a claim about who work belongs to.

What I want from the next workspace is more human visibility, not less: people seeing each other's work, and each other's agents' work, and spending their attention on judgment instead of transport. A team has always been the technology that makes individual intelligence compound. That property is worth carrying through this transition, and it will not carry itself.