Today’s software is organized around applications and artifacts. Documents, tasks, messages and spreadsheets became the unit because computers could not model the work itself. The work gets translated into whatever abstractions each application understands, and people spend a second layer of effort keeping those abstractions synchronized as reality changes.

I came at this first as an attempt to build a genuinely persistent, always-on assistant. The limiting problem turned out not to be the interface, or another agent framework. It was that the system had no durable representation of the world and the work it was supposed to reason about.

WorkspaceOS starts from a more fundamental primitive: the project, defined as any directed effort. It maintains a persistent model of that effort, holding its entities, claims, decisions, dependencies, actors, goals and changing state. Documents, plans and tasks become projections of that model rather than separate systems of record.

The model stays grounded because every claim traces back to its source. New information is classified as novel, confirming, redundant or contradictory, and where evidence genuinely conflicts the model represents the claim as contested rather than quietly picking a winner. Documents generated from it carry statements linked to the material they came from.

It is operational rather than archival. Correcting a statement in a generated document updates the underlying claim and propagates through everything derived from it. A change in the model can prompt the planner to revise its plan without being asked.

That makes a different class of software possible: software that notices when reality changes, understands what that change affects, and moves the work forward without asking you to reconcile everything by hand. The open question is no longer whether the model can be built. It is where it creates enough leverage that the way you work actually changes.