OOmnistu

Blog / AI Development

AI Development

Android's agent playbook: Android MCP, the Agentic API and the Connected Apps waitlist

Google's 5 November 2026 Android Developer Summit post names the integration paths: Android MCP on AppFunctions for agent tool-calling, ML Kit GenAI's Agentic API on Gemini Nano 4, the A2UI Compose renderer in alpha, and the gated Connected Apps Platform - with Google's benchmarks labeled.

OOmnistu
Reviewed 2026-11-19 · 4 min read

Google used the Android Developer Summit on 5 November 2026 to say plainly what the last two years hinted at: on Android, the user of your app is increasingly an AI agent acting on a person's behalf. The follow-up blog post lays out the integration paths, and for once they are concrete enough to plan against. Here is what was actually announced and what an app team should do with it - with Google's performance claims labeled as Google's.

Path one: let agents call your app - Android MCP and AppFunctions

The headline mechanism is Android MCP, a tool-calling framework built on the MCP standard and powered by the AppFunctions Jetpack library. It exposes your app's capabilities as functions that on-device agents like Gemini can invoke. Google positions it against a remote MCP server: remote gives you cross-platform reach, while Android MCP is the recommended path when the journey touches end-to-end-encrypted or sensitive local data, needs device sensors or hardware, must execute locally, or has no server endpoint at all. You can ship either or both. When no integration exists, agents fall back to Android's Computer Control framework to drive your UI directly - which is slower and more brittle, and is Google's argument for doing the explicit integration.

The framing around this is the "Android intelligence system": prompted assistance (the user asks an agent to chain tasks across apps) and proactive assistance (the system surfaces suggestion chips on lock screens, notifications, keyboards and inside apps). Your functions become the verbs in both modes.

Path two: put the agent inside your app - ML Kit GenAI's Agentic API

On the in-app side, ML Kit GenAI APIs now include an Agentic API alongside the existing Prompt API: it auto-executes function calls, keeps conversation history, runs the orchestration loop and applies automatic guardrails, powered on-device by Gemini Nano 4 - which adds function calling. Google claims Nano 4 delivers up to 3x inference speedup and 4x lower memory use than the previous Nano under its own internal testing benchmarks; treat those as vendor numbers until independent devices confirm them. Around it: an alpha Embeddings API, Prompt API additions like thinking mode and system instructions, and LiteRT in Google Play services for custom models - which moves the runtime out of your APK (Google says about 8MB smaller) and now covers Wear OS and Automotive. For hybrid workloads, Firebase AI Logic routes between cloud Gemini models and on-device Nano, with App Check, server prompt templates and spend caps for production control.

Path three: let the agent draw - A2UI Compose renderer

The Compose A2UI renderer, now in alpha, lets an agent stream interactive native Compose UI from a JSON protocol instead of replying in prose - the keynote demo was a booking assistant rendering its own interface. If you have watched agents fumble form-filling in chat bubbles, the appeal is obvious; the alpha label is the caveat.

The distribution piece: Connected Apps Platform

Tying it together is the Connected Apps Platform, Google's name for reaching users across the Gemini app and Google Search on Android: Android MCP for on-device execution plus remote MCP for reach. You can build and test locally today with Gemini Apps Developer Mode, but production access is a preview waitlist, currently limited to selected production-ready apps in supported categories, with broader self-serve promised later. So the discovery upside is real but gated - this is a join-the-waitlist item, not a ship-this-quarter item.

What to actually do

The honest summary

Android's agent story is no longer a keynote concept: there is a named tool-calling framework (Android MCP on AppFunctions), an in-app agent API with guardrails (ML Kit GenAI Agentic), a UI-streaming renderer in alpha (A2UI Compose), and a gated distribution channel (Connected Apps Platform). The integration paths are documented and testable locally today; the performance numbers are Google's own, and the widest distribution sits behind a waitlist. The useful move this quarter is small: name your app's verbs, prototype one as an AppFunction, and get on the list.

Sources

Prepared with AI assistance and reviewed before publication. Framework names, integration paths and preview status are drawn from the source post above, checked on 19 November 2026; benchmark figures in it are Google's internal testing claims and are attributed as such.