Is Android ready for AI agents? What broke in 3324 actions on 45 phones

The hard part of letting an AI assistant use an Android phone turned out to be reaching the phone, not reading its screen. Opposable, an Android app that lets Claude, ChatGPT, Claude Code or Codex operate the phone it is installed on, looked at 3,324 actions that assistants made on 45 phones between 8 October and 3 November (Opposable's analysis). 91 percent worked, and taps and swipes did what the agent asked 99 percent of the time. The most common failure was a phone that was no longer connected: 27 of the 45 phones dropped off Opposable's relay while an agent was partway through using them, accounting for 108 of the 305 failed actions.

The cause is how Android saves battery. Features called Doze and App Standby defer background work on a phone nobody is touching, and several manufacturers add battery managers that close apps Android itself would leave running. An agent usually arrives at exactly that moment, when the owner has asked from a laptop or put the phone in a pocket, and if the app's connection to the relay is closed the phone stops answering. 36 of 105 logged sessions hit a disconnected phone, and the agent gave up in 16. Seven of nine Xiaomi phones and 14 of 20 Samsung phones dropped off at least once, against two of five Pixels and one of four Motorolas. The log cannot separate a battery manager closing the app from an owner switching the phone off or losing signal, but the split by maker matches dontkillmyapp.com's ranking of aggressive battery managers. Opposable now runs as a foreground service with a permanent notification, asks to be exempted from battery optimisation and sets a 15-minute reconnect alarm, and the author notes that not every manufacturer respects the exemption.

The second limit never shows up as a failure. Agents act through Android's accessibility service, which was built for people who use screen readers or switch access, and to Android those taps look exactly like the owner's. No flag marks input as coming from an agent, so apps cannot treat it differently, and the owner has no system setting that keeps an agent out of one app or asks before it acts. Android 16's AppFunctions lets an app offer actions to agents directly, but only where developers adopt it. Opposable built its own checks instead: payments wait for the owner's approval on the phone, and the app's own screens refuse agent taps, which happened four times in the period. The author asks Android for a supported way to stay reachable that manufacturers respect, and for a notion of agent input with per-app permission, a marker on each event and a prompt before sensitive actions.

Smaller problems were fixable by the app. Typing worked 85 percent of the time, failing 21 times because text arrived before a field had focus, and opening an app failed 15 percent of the time partly because the same app has different package names on different brands. The 99 percent tap figure also flatters coordinate taps, which always land somewhere, compared with 92 percent for taps on a named screen element. These are Opposable's own logs for its own product, from 45 phones over four weeks, and 3,048 of the actions came through Claude, so the results mostly describe Claude's habits. The log keeps 200 entries per phone, so the earliest sessions on the five busiest phones are missing.