Product
Android 1.20: the agent operates your apps without missing a tap
We ran the same 20 phone tasks until two runs in a row went through with no failed tap and under 10 seconds of tool time each. The APK downloads now require a signed-in account.
How we measured it
We gave the agent the same 20 tasks again and again on a Nothing Phone (3a) Pro: a note in Keep, a contact, a calendar entry with date and time, a translation, a bike route in Maps, searches in DuckDuckGo, Firefox, YouTube Music and the Play Store, and a few edits and deletions. After every run we read the full log of each tap and fixed what went wrong. A version counted as done only when two runs in a row had no failed tap and every task needed less than 10 seconds of tool time. Model thinking time is not included. Version 1.20.0 is the first to pass, with 9.6 and 9.4 seconds for the slowest task.
What changed in the tools
Most fixes were about waiting for the right moment. The agent used to read a page before it had loaded, or read a list while a menu was still closing, and then tapped something that was gone a moment later. It now waits for the content, and it notices when a map keeps redrawing but nothing on it changes. Suggestions that load after typing, such as place names, are now in the list before the agent picks one. A time picker showing the same hour twice no longer confuses it. None of this knows a particular app, so it applies to any app the agent operates.
Downloads need an account
The APK files are now only handed out to signed-in accounts. Before today anyone with the address could download them, and a copy could stay in the cache of our network provider for hours. If you are not signed in, the download button takes you to the sign-in page and back to the app page. The app is useless without an account anyway, so nothing changes for people who use it.