This is the right framing. For agents, the hard part isnât âone-click task completionâ, itâs a reliable control loop: observe the real device, take a small deterministic action, read the result, then repeat.
Screenshots, tap/swipe/text/keyevent/shell/install, Relay routing, request ID matching, and execution queues may look like low-level infrastructure, but theyâre exactly what makes long-running Android automation dependable.