In my opinion the OS should broker the communication between harnesses (which is part of Siri), LLMs and app functionality. It can take into account the permissions, the user could configure which LLM to use at an OS system settings level and complex workflows can integrate permissions like FaceID only once as the harness then interacts with APIs exposed by applications (arguably already there with app intents).

Self promotion blog article:

https://braw.dev/blog/2026-04-06-mobile-os-as-agentic-broker...

Feels like this is based on the idea that any model is interchangeable with any other, while all of them are better at some stuff, worse at some stuff.

Besides, the modern concept of "agent" is what, some year old? Bit fast to ossify stuff in the OSes already.