OS Level Control for AI Agents

This is why the real fight in AI agents is moving below the app layer and toward control of the operating system. A normal app can only see its own little box, or whatever other apps choose to expose through APIs like AppIntents. An OS can watch the full sequence of what someone opens, reads, compares, and does, which is what turns an assistant from a chatbot into software that can actually anticipate and coordinate across apps.

• Launchers, assistant replacements, and accessibility hacks each unlock a slice of context, but none give the same system wide visibility as forking Android. A launcher can see what app was opened. Accessibility can read what is on screen. The OS can combine those signals with notifications, history, and actions into one model of intent.
  1 sacra
• Perplexity shows the ceiling of the app and assistant approach. Its newer push is into an agentic browser, because search with citations inside an app was easy for model labs to copy, while owning the interface where actions start gives deeper context and more room to execute tasks.
  2 sacra
• This also changes what apps become. Instead of being the main place where users browse and decide, apps start looking more like back ends that hold inventory, identity, payments, and data, while the OS or agent decides when to call them and what to surface.
  1 sacra 3 sacra

Over the next few years, the winning AI interface is likely to be the one that sees the most context with the least friction. That points toward browsers, OS level assistants, and new AI native devices converging on the same role, a layer above apps that reads across systems and routes work to them in the background.