For fifteen years we've designed Android apps around one assumption: a human will tap the screen. App Functions quietly break it. With the latest platform, an on-device agent — Gemini — can now discover your app's capabilities and invoke them directly, no UI in the loop.

This talk is not a feature tour. We start with the mechanics (AppFunctionService, schema metadata, ExecuteAppFunctionRequest, URI grants), then spend most of our time on the part the docs don't tell you: the design shift. When your business logic becomes a public, schema'd tool surface, hard questions appear. Which capabilities do you expose — and which do you never hand to a model? How do you make functions idempotent and safe when invocation is non-deterministic? What does "offline-first" mean when an agent, not a user, is driving a stateful app? And how on earth do you test something an LLM decides to call?

You'll leave able to ship your first App Function — and, more importantly, with a mental model for treating agents as first-class callers of your app, not a gimmick bolted on top.