Today's phone-navigation agents are either too expensive to use or trained on static examples, making them brittle when they hit an unexpected screen.
This work builds a two-part system: a planner that chooses actions and an execution model that taps the screen. The planner learns online on real iPhones using reinforcement learning, guided by a separate judge AI that scores every single step instead of just the final outcome.
The resulting 8-billion parameter agent matched the performance of the most expensive commercial models on real apps and was seven times better than standard methods on Android devices it had never seen, proving online step-level training creates robust generalizable agents.