An AI assistant app on your phone can answer almost any question but do almost none of your tasks. Ask it anything and it answers. Ask it to book the cab or clear your inbox and it stalls. The reason isn't that the model isn't smart enough. It's that it isn't allowed to. To act instead of just talk, an assistant needs to do things a normal app can't. Hear you without being opened first, see what's on your screen, reach into your other apps, keep running in the background. On Android, only Gemini ever got those privileges. Any other assistant you download runs as a boxed-in app, able to chat but not to act. On July 16, though, the EU ordered @Google to change that, handing 11 of these privileges to rival assistants or risking a DMA fine of up to 10% of global revenue. These privledges sort into four buckets: 1️⃣ Getting summoned: waking by voice, launching from the home button 2️⃣ Reading your context: seeing your screen, sensors, and app data, offering things before you ask 3️⃣ Acting for you: running tasks inside your apps, automating multi-step jobs, changing settings 4️⃣ Running the engine: using the phone's built-in models, running its own, staying alive in the background However, Gemini was the assistant Google built and tested for its own OS. Swap in a different one and you're trusting a new model (and a new runtime) with permission to act on your screen and inside your apps. A weaker or lower-reasoning model driving your OS doesn't just answer worse, it acts worse, and a wrong action costs more than a wrong sentence or hallucination. You can't swap the model under an agent and expect the same results, let alone the same safety. So opening the layer solves one problem and surfaces a bigger one. Android is only the first place this shows up. The same fight is already moving to popular browser, desktop, and car OSs where there's even more potential for major vulnerability issues. As full trustless agent automonomy is still quite a ways away, it's interesting to see how quickly adoption is growing while full security and trust still lag.(Sahara AI 🔆)