OS Sandboxing Restricts AI Assistants
Sam Hall, CEO of Wafer, on AI agent form factors
The real constraint is that Apple and Google sit in the middle of a contract heavy marketplace, not a blank technical canvas. Their operating systems deliberately sandbox apps, and their assistant frameworks only let apps expose specific actions they choose to expose. That keeps the platform legally cleaner, but it also means any assistant built by the platform owner is boxed into narrow, developer approved workflows instead of freely reading across Uber, Lyft, Gmail, Outlook, and every other app on the phone.
-
On Android and iOS alike, the official assistant path is opt in. Google App Actions require developers to declare capabilities in shortcuts.xml and get review before users can access them. Wafer describes Apple and Google inventing AppIntents and similar hooks because the assistant cannot simply inspect app data by default.
-
That limitation is strategic, not just technical. Google makes money from Google Play, and Wafer frames Android as a distribution layer for the store. If the OS starts completing tasks without opening apps, apps become back end services and the app store loses some of its control over discovery, engagement, and payments.
-
The best comparison is Granola. It wins by sitting closer to the microphone and calendar than Zoom does, so it captures the raw meeting workflow before Zoom can package it. Wafer is making the same bet at the phone level, that the company controlling the read layer can reshape where value accrues.
The next step is a split market. Big platform owners will keep expanding safe, permissioned assistant rails, while startups and OEM aligned Android forks push deeper OS level access to build more capable agents. If those products create real user pull, app developers will gradually adapt and ship toward an agent first world where apps are increasingly infrastructure, not destinations.