System Access Over Model Quality
Sam Hall, CEO of Wafer, on AI agent form factors
The defensible layer in consumer agents is shifting from model quality to system access and action reliability. A lab with the best model can still feel shallow if it only sees what a normal app is allowed to see. Wafer is betting that the winning phone agent will come from owning the layer that can read cross app context, watch repeated user workflows, and trigger actions with far less handoff friction than an app or assistant API allows.
-
The core limitation is Android sandboxing. A normal app can only access data another app chooses to expose, so it cannot compare Uber against Lyft prices, learn broad usage patterns, or automate a flow it cannot observe. That makes OS control a product capability, not just an engineering preference.
-
Perplexity shows the middle path. By becoming the default Android assistant, it gets a privileged entry point and can call app provided intents, but those actions are only as rich as developers make them. That creates a narrower experience than an OS fork, even with strong models and distribution.
-
This mirrors other AI products moving lower in the stack. Granola improved meeting notes by listening at the microphone layer instead of waiting for Zoom APIs. The pattern is that better context capture often beats better raw model quality, because the model cannot reason over data it never sees.
The next phase of competition is likely to split in two. Frontier labs will keep pushing broad assistants through apps, browsers, and default surfaces, while startups and OEM aligned players push deeper into launchers, assistants, and Android forks. If agents become the main way people use phones, the companies that control system context will shape the user experience above the model layer.