Still a work-in-progress[1], but I have a mobile app for herdr also built with a native GPU interface (but not using gpui), optional LLM usage (outside of the agents themselves), designed for a hands off voice-enabled approach to development. The STT/TTS are both on-device. Yes, I used agents to build it, and you can see the development record in tree under docs/design.
[1] https://github.com/flankrapp/flankr