Gemini Spark code reveals skill system

Google’s emerging Android AI stack is starting to show what a consumer agent may actually run on. Reports published on May 14 and May 16 pointed to code in the Google app and Android 17 that references “Gemini Spark,” a product name tied to an agent flow with skills, scheduled tasks and background execution. Google has not publicly detailed Spark in full, but its developer documentation now describes adjacent pieces — AppFunctions, Android Computer Control and AICore — that match the architecture those reports describe. The result is a clearer picture of where Google is heading on Android. Instead of relying only on a chatbot interface, the company is laying groundwork for software that can discover app capabilities, invoke them locally and carry out multi-step actions on a phone under system controls. Google’s public materials stop short of naming Spark as a launched product, and the company says some of the enabling pieces are still limited to previews and trusted testers. (9to5google.com) ### Where does the “Gemini Spark” name come from? 9to5Google reported on May 14 that Google app beta version 17.23 introduced the “Gemini Spark” name after earlier references to “Gemini Agent.” The publication said Spark appeared in a redesigned Gemini app navigation flow with separate “Chat” and “Agent” tabs, plus lists for active and scheduled tasks. 9to5Google described the findings as an APK teardown, noting that such features may not ship and that interpretations of code can be imperfect. (developer.android.com) Forbes reported on May 16 that hidden Android 17 code pointed to a broader Spark system with a skills registry and task scheduler. That report described an always-running background service and device checks that could limit availability based on hardware and permissions. Forbes’ account is based on code analysis rather than a formal Google announcement. (9to5google.com) ### What public Google technology matches those reported pieces? Google’s Android developer site says Android is becoming an “intelligence system” and tells developers to expose app capabilities to “qualified agents” through AppFunctions, or fall back to UI automation on supported devices. The same page says AppFunctions integration with Gemini is in a private preview with trusted testers. (forbes.com) Google says AppFunctions is an Android API that lets apps expose specific functions for callers such as agent apps to access and execute on device. In a February explanation of the feature, Google compared AppFunctions to the Model Context Protocol, but said the functions run locally on Android hardware. Google’s examples include creating calendar events, finding information in one app and passing it into another, and building cross-app workflows from a single request. (developer.android.com) ### How would skills fit into that system? Google’s developer materials already use the term “skills” in its coding tools. Android Developers says skills are AI-optimized instructions that help agents understand and execute specific patterns, and that relevant skills can be pulled in dynamically when needed. Those pages are aimed at Android Studio and Android CLI, not the consumer Gemini app, but they establish Google’s own vocabulary for modular agent capabilities. (9to5google.com) The Android side of the stack uses a different public noun for app actions. Google says AppFunctions lets apps behave like on-device MCP servers, contributing functions that assistants can call as tools. Taken together, the public documentation and the reported code suggest a split between a user-facing agent layer and a lower-level capability layer exposed by apps and system services. That reading is an inference from Google’s documents and the reported teardowns, not a statement Google has made directly about Spark. (developer.android.com) ### Why do device requirements matter so much here? Google says Gemini Nano runs in Android’s AICore system service, which uses device hardware to enable low-latency on-device inference and local data processing. Google first positioned AICore as available on select devices, starting with Pixel 8 Pro, and its newer materials point developers to previews for Gemini Nano 4 on future flagship Android devices. (developer.android.com) Android Computer Control adds another limit. Google says that framework allows OEM-preloaded AI assistants to automate tasks in selected apps installed on the device, and specifically says it is for OEM-preloaded assistants on supported devices. That means broad agent behavior on Android is not just a model question; it depends on which phones ship the right system services, which assistants are preloaded, and which apps expose callable functions or support controlled automation. (developer.android.com) ### What has Google actually said about safety and permissions? Google’s public Android guidance says developers can dynamically enable or disable AppFunctions so the system knows which features are available for a user at any given time. That documentation frames gating around account state and availability rather than unrestricted access. (developer.android.com) 9to5Google also reported that Spark’s in-app disclosures say the feature is experimental, may ask permission before sensitive actions, and could share information with third parties when needed to complete a task. Those strings were found in app code, not in a finalized product policy page, but they indicate Google is preparing user-consent language around purchases, personal data and supervised use. (developer.android.com) ### What should readers watch next? Google I/O 2026 is the next obvious checkpoint for formal product details, and Google has already used recent Android developer posts to introduce “Gemini Intelligence,” AppFunctions, Android Computer Control and AICore-related previews. As of May 2026, the Android Developers site says AppFunctions integration with Gemini is still in private preview with trusted testers, while Gemini Nano 4-enabled devices are expected later this year. (9to5google.com) (developer.android.com)