google i/o keynote wrapped a few hours ago. the headline was gemini spark — the new agent layer for android 17, integrated into the gemini app, capable of running multi-step tasks across gmail, docs, calendar, and any third-party app that exposes the right surface.
most coverage today will frame this as "google's answer to apple intelligence." that framing misses it. the part that matters for indie devs isn't the agent. it's the skill system sitting underneath it.
what was actually announced#
gemini spark isn't one product. it's three:
- the agent — gemini decides what to do, calls the right tools, executes a plan
- the skill system — users author named, reusable workflows like "summarise my morning emails into a slack post"
- the task scheduler — skills can run on a schedule, not just on demand
the agent gets the demos. the skill system gets the platform.
why the skill system is the story#
an app actions integration in the old android intents world was a one-shot. a user said something, gemini parsed it, your app got an intent, your app responded. no memory, no chaining, no scheduling. a smart voice command, not an agent surface.
the skill system changes that. once your app exposes an intent gemini can call, it becomes a building block any user's skill can reference. a recipe app's "add to shopping list" gets composed into a saturday meal-prep skill. a fitness app's "log workout" gets called automatically at the end of every run. you don't write the skill — the user does. you just have to be reachable.
apps that ship rich, well-named app actions become lego pieces in other people's workflows. apps that don't, become apps the agent has to open and tap through pixel-by-pixel — the slow, fragile path google is quietly deprecating.
why this matters for indie devs#
new surfaces always reward early movers, but the reward this time is structural. if spark's skill ecosystem behaves anything like shortcuts on ios or zapier on the web, the apps that show up in the autocomplete when a user types "add to my skill…" are the apps that get used. the apps that don't are invisible — reachable only by opening the app drawer, which is exactly the behaviour spark is built to replace.
the work to be on that list isn't huge. mostly:
- declare your app actions in your manifest with names a human would use
- make sure each action can be called headlessly, without a forced ui
- write descriptions that match the verbs users will say to their agent
- list those capabilities clearly on your play store page, so the agent's app discovery layer can find them
the stora angle#
the third item — descriptions and capabilities on the play store page — drifts every release. your build adds three new app actions; your listing still describes the app it was a year ago. that's the gap stora closes: it reads your manifest, finds the app actions you actually ship, and keeps your store listing in sync with the verbs your app responds to. when gemini's discovery layer indexes the play store for a "log a run" action, your app shows up because the listing tells the truth about what your app does.
the call#
watch the developer keynote later today for app actions specifics. then open your manifest and count the intents. if the answer is "two and i wrote them in 2023," that's the work. agent platforms reward apps with surfaces and punish apps that hide behind a launcher icon.
spark ships in stages over the summer. the apps on the skill autocomplete on day one are the apps that win.