Three feature launches from the past six weeks that I had been reading as unrelated, and they aren't.
Cursor 3's event-triggered automations: the agent acts on conditions, not on invocation. MDASH: adversarial agents debate, posterior-credibility scoring routes uncertain findings to a human. Claude Code's Dreaming feature: the agent reviews its own logs between sessions and self-corrects offline.
Three vendors, three surfaces. What they share is that none assumes a human watching the trajectory in real time. They are three sketches of what oversight looks like after you accept that supervisory monitoring — human attention proportional to agent activity — does not scale to the volume agents are now generating.
The sketches answer the bottleneck differently. Cursor moves the consent moment from invocation to trigger specification — you authorize the condition. MDASH replaces continuous attention with disagreement-routed attention. Dreaming moves the correction loop inside the agent — the failure is adjusted before it surfaces.
The supervisory framing felt novel two years ago. It now looks like a transitional pattern. The next architecture will be one of these three shapes, or a composition — and the version that lasts will be the one that handles the failure case all three currently dodge: when the system itself is the thing drifting, and no human is in the loop close enough to notice.