Developer's 18 months with Android's agentic APIs exposes what's real and what's provisional

One broken alpha sets the tone

A developer who has spent a year and a half building against Android's agentic APIs has published a first-hand account on dev.to that takes a skeptical look at Google's framing of Android as an "intelligent system" rather than a conventional operating system. The author's app, Be nice, gained an AppFunction so Gemini could invoke its createAppPair capability directly, and the integration later broke when a library bump from alpha08 to alpha10 quietly dropped a service component the earlier alpha had shipped; the fix was to host that component manually. That experience, the author writes, is a useful filter for keynote language.

According to the dev.to post, the "intelligent system" label covers three separate engineering questions: the on-device model runtime and the stability of its APIs; app integration, meaning how apps expose functionality to agents and who controls that exposure; and distribution, meaning which assistant gets to stand in front of the user. Google's messaging merges all three. A February 2026 "Intelligent OS" post redefined app success as completed tasks rather than app opens, and the May I/O recap claimed that agents such as Gemini can already navigate an app on a user's behalf, with no one tapping through the interface.

Distribution is the weakest layer

The author starts with distribution because the marketing runs furthest ahead of it there. Android Halo, a status-bar slot that surfaces agent activity as it works, ships with Android 17 later in 2026. In a July interview, Android president Sameer Samat described the slot as belonging to whichever agent a person chooses, Gemini or otherwise. The post counters that a status-bar slot is not an API: genuine parity would require the same wake-word rights, background-execution allowances, on-device context access, and permission for integrated actions that Google's own assistant gets by default.

The verification mechanism arrives later, and only in Europe. On July 16, 2026, the European Commission adopted a Digital Markets Act decision ordering Alphabet to give third-party AI assistants eleven Android capabilities at no cost and on terms as effective as those available to Google, grouped into invocation, context, actions, and on-device resources. Most of it must ship with Android 18 by August 1, 2027, with concurrent wake-word detection deferred to Android 19 by August 1, 2028. Until then, outside the EU, Halo's openness rests on an interview quote that the author considers revocable. The post also recalls the earlier browser choice-screen remedies, which Mozilla described as uneven across the EEA with the clearest gains coming on iOS, as a warning that access mandates rarely dislodge a pre-installed, pre-authorized default.

AppFunctions is genuine, and still moving

At the app-integration layer the author is far more positive, calling the architecture real. An app extends AppFunctionService, annotates callable methods and their parameter and return types, and an annotation processor generates an XML schema that the OS indexes. The schema even carries the developer's documentation, which serves as the natural-language contract the agent reasons over. Agents discover what a package offers through AppFunctionManager, gated by the EXECUTE_APP_FUNCTIONS permission. The post likens the setup to an on-device MCP tool server, except the platform itself is the transport and the "server" is an app the user already installed.

The gaps, though, are structural rather than preview-stage rough edges. The permission is one blanket grant rather than a per-function one, the platform does no rate limiting, and validation stops at the type level: a call that is structurally valid but semantically nonsensical, such as booking a ride in the past, sails through because business rules remain the developer's job. The safety model is to notify the user and let them intervene, the same posture MCP itself is criticized for. AppFunctions requires Android 16 (API 36), and Gemini's side of the integration was still a private, trusted-tester preview as of May 2026. Samsung is the standout production reference point, shipping Calendar, Notes and Tasks AppFunctions on the Galaxy S26 under OneUI 8.5, and Google previewed a rougher second path at I/O in which an agent drives ordinary apps through UI automation. The author's advice to other developers is blunt: treat every surface in this layer as provisional and budget for at least one more breaking rename.

Why it matters

If completed tasks, rather than app opens, become the measure of an app, the shift Google's February post anticipates, then the stability of AppFunctions, the neutrality of Halo, and the DMA compliance schedule together decide whether Android's AI era is an open platform or a single-vendor channel with regulated edges. The dev.to account matters because it comes from someone who shipped against the stack: it separates what demonstrably works, such as the AppFunctions mechanism and Samsung's deployment, from what is still a promise, such as third-agent parity and API stability. For teams planning agent-facing features now, the practical takeaway is to build, but to keep a breaking-change budget on the roadmap.