OS-Level AI Agents Predict Intent
Sam Hall, CEO of Wafer, on AI agent form factors
This points to the real prize in AI assistants, which is not better chat, but privileged access to the user’s full context. Wafer’s argument is that once software can see calendar events, app usage patterns, conversations, and repeated actions across the phone, it can move from waiting for commands to surfacing likely next steps, like checking Uber and Lyft before a meeting. That is why companies are pushing below the app layer and closer to the operating system.
-
The core divide is read first, write second. Wafer describes the hard part as building enough context to understand the user, then using that context to trigger actions. It trains on repeated behaviors, like watching someone open Spotify, search an artist, and press play, then reusing that pattern later with different inputs.
-
App level assistants are narrower because apps are sandboxed. Perplexity Assistant can act through developer exposed AppIntents, but that only covers actions an app explicitly allows. A forked OS can compare ride prices across Uber and Lyft or connect signals across apps because it sits above those silos, not inside one of them.
-
This is the same product move seen in other AI software going lower in the stack for richer data. Granola watches mic activity, calendar invites, and meeting URLs at the desktop layer so it can infer a meeting has started and launch note capture automatically, instead of waiting for a user to prompt it inside one app.
The next step is a shift from apps as destinations to apps as back end services. If OS level agents keep getting better at predicting intent from behavior, the winning mobile experience will look less like tapping through icons and more like approving, editing, or declining suggested actions assembled from many apps behind the scenes.