Most agent architectures are stuck in the uncanny valley of simulation: they pretend to be a human clicking buttons, even when they could just talk to the system directly. Aico’s latest architecture for Phone Bridge breaks this loop by introducing a hybrid routing mechanism that prioritizes software fast paths over hardware emulation. The default state remains USB HID for input and HDMI for screen observation, making the host see a standard keyboard and mouse. But when a task demands data retrieval—like checking a calendar or reading contacts—the system checks if a companion app can handle it via API before falling back to the slow, error-prone visual navigation.

The Routing Logic: API First, Pixels Second

The core innovation here is the decision tree at the entry point of any agent action. The system asks a simple question: Is a software fast path available? If the answer is yes, the action is routed to the companion app’s API. If not, it goes down the hardware path, relying on HDMI observation and HID input. The companion app exposes specific operations such as launch_app, clipboard_read, clipboard_write, calendar_read, contacts_read, and notifications_read. This isn't just convenience; it's a reliability strategy. By bypassing the UI for structured data, Aico eliminates the failure points inherent in visual models, such as unexpected modals or loading states.

Platform-Specific Optimizations

The routing logic adapts to the constraints of mobile operating systems. On iOS, Aico leverages Dynamic Island and Live Activities to fetch task state over USB ECM, avoiding cloud latency entirely. Picture-in-Picture (PiP) allows the agent to remain active even when the app is backgrounded, while BLE Wake ensures idle devices can be brought up for immediate use. Android gets similar treatment with a foreground service for continuity and a companion app that mirrors iOS operations. The key takeaway is that the hardware fallback is reserved strictly for operations that genuinely require visual context, ensuring the agent isn't wasting cycles simulating clicks when it can just ask for the data.

Key Takeaways

  • Hybrid routing reduces failure points by skipping UI navigation for structured data tasks.
  • Companion app APIs handle clipboard, calendar, and contacts operations directly.
  • Hardware control (USB HID/HDMI) remains the universal fallback for visual-only tasks.
  • iOS optimizations include USB ECM for Live Activities and BLE Wake for device activation.

The Bottom Line

This is how you stop building brittle demos and start shipping reliable agents. If you're still forcing LLMs to look at pixels to read a calendar, you're doing it wrong.