The thesis is that a longer-running agent becomes a product-runtime question: its work surface, context handling, and review path matter alongside the model.
The mechanism is an agent harness with a dedicated computer and browser, alongside improved context management and a named review mode. The public description does not explain how the computer is isolated, what permissions agents receive, or whether safety checks block actions pending approval. Why now: Notion is explicitly extending agent work from shorter interactions to runs lasting as long as a day, making supervision and continuity consequential product features rather than background implementation details.
The first implication is that teams evaluating these agents should assess the runtime and review controls, not just task quality or duration. The first watch condition is whether Notion publishes details on isolation, permissions, and the points at which human review is required. These are implications to test, not confirmed properties of the system. Evidence boundary: the release page states the features and duration but does not provide technical architecture, independent evaluations, or operational outcomes. This starts a distinct analytical record because no prior New Runtime record was supplied; no earlier record is being assumed.