Aiden describes an architecture for agents that requires full-duplex voice conversation alongside complex visual reasoning and device actions, avoiding the limitation of a single model handling both. The solution splits responsibilities: a real-time voice model manages the conversation while a separate, stronger model executes background tasks, coordinating asynchronously via a task queue.
Key design decisions include tracking task completion and success as distinct signals to prevent the conversational layer from guessing backend outcomes. Results reaching the foreground pass through a queue with a 500ms aggregation window to merge concurrent outputs. Device tasks are executed strictly serially to avoid ambiguity, and runtime events are converted to standard UserMessage objects to avoid frequent system prompt rewrites.
The author notes this is a dev-board stage implementation and seeks feedback on interruption handling when new voice input arrives while the backend is mid-task.