Skip to content

Build1 publisher3 min readPublished

Bumping one alpha version broke the AppFunction that let Gemini call createAppPair

A developer who exposed createAppPair to Gemini lost the integration to a library bump. The same post argues Android Halo ships about a year before anything exists to check its neutrality promise against.

The Engineer · Build desk

Illustration accompanying Bumping one alpha version broke the AppFunction that let Gemini call createAppPair

What happened

  • An AppFunction that let Gemini invoke createAppPair in the app Be nice stopped working after a single alpha bump, because alpha10 no longer shipped the service component alpha08 depended on.
  • Android Halo, a new status-bar slot that shows an agent's status while it works, ships with Android 17 later this year.
  • The European Commission adopted a decision on July 16, 2026 under Article 6(7) of the DMA ordering Alphabet to give third-party AI assistants eleven Android features free of charge on equally effective terms.
  • Most of those features are due with Android 18 by August 1, 2027, with concurrent hotword detection deferred to Android 19 by August 1, 2028.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

  • constraint While the integration libraries are in alpha, a working agent contract can be removed by the dependency rather than by the developer, so every bump means re-verifying the invocation path along with the build.
  • decision Teams budgeting agent work have to choose a layer to spend on, and the surface Google's task-completion metric depends on is the one the post trusts less than the runtime.
  • precedent The Commission's four buckets give agent parity a checkable definition, so any vendor claim of neutrality can be tested against invocation, context, actions and resources even outside the EU.
  • contradiction Halo arrives on a stated commitment to whichever agent the user picks, while the features that would let a rival assistant test that commitment are due roughly a year later and only in one jurisdiction.

An AppFunction is two things: a declaration inside your app, and a delivery path the library provides. The declaration is yours. The service component that carried Gemini's call into Be nice came from the library, and when the newer alpha stopped shipping it, the integration went with it [1]. Anything you expose to an agent through an alpha library lasts as long as the library keeps shipping the parts underneath it.

The post splits Android's "Intelligence System" into three layers: the model runtime, app integration, and distribution [4]. It says only one of the three is finished enough to build on [5], and it takes them in order of least trust first, starting with distribution [6]. Read against the numbering, that puts the runtime at the top and integration in the middle [23]. The author has been building against these APIs for the last year and a half [2], and the one failure he documents is in the middle layer [1].

That middle layer is where Google's own success metric now lives. The February 2026 "Intelligent OS" post redefined what a successful app looks like as completed tasks, with or without a human touching the UI [7]. The May I/O recap said agents like Gemini can already "complete task automation" on Android, navigating an app on the user's behalf [8].

Distribution is the layer with the least shipping code behind it. Google describes Android Halo as making "your agent's status visible by bringing subtle communication to the top of your phone screen as it takes on a task, goes into live mode or sends you a message" [9]. In a July interview, Sameer Samat, Android's president, described the slot as agent-agnostic, a dedicated spot for whichever agent a person chooses, Gemini or otherwise [11]. What exists so far is a status-bar slot and an interview [12].

The enumerated list comes from Brussels. The eleven features in the Commission's decision group into invocation, context, actions and resources, holding two, three, three and three items, which is where the eleven comes from [15][19]. Concurrent hotword detection, the one that lets a second assistant listen for its own wake word at all, sits twelve months behind the rest [20].

Set the developer's test beside the remedy and they ask for the same things. The post's minimum for a non-Gemini assistant on equal footing is four items: the same wake-word rights, the same background-execution allowance, the same access to on-device context, and the same permission to perform integrated actions such as sending an email or booking a ride [13]. Each has a home in the Commission's buckets, with wake word under invocation and background execution under resources [22]. Halo arrives with the promise; the checkable version arrives roughly a year later, inside the EU [21]. Elsewhere, the post argues, that openness is a line from an interview, revocable, on a schedule Google sets, on a platform Google also uses to distribute its own competing product [18].

For the alpha10 number to transfer to your app, you need the same shape of dependency: an agent-facing function whose invocation path is owned by a library still in alpha, and a build that tracks alpha versions. This is a narrow claim about one integration in one app, and the only measured evidence in the piece.

What to watch

  • Whether AppFunctions leave alpha with a stability guarantee for the invocation path, not just the declaration API.
  • Whether Android 17's Halo documentation names an interface a non-Gemini agent can use to occupy the slot.
  • Whether Alphabet's Android 18 implementation lands the eleven features by August 1, 2027, or the deadline moves.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories