AI coding agents increasingly run tasks that outlast a single desktop interaction. That creates a genuine interface problem: how can a developer see progress, respond to a safe prompt, or stop a task while away from the keyboard—without turning convenience into uncontrolled remote access?
AtlasRepo tested an open-source Apple Watch bridge in both a short real-world demo and a full architecture and setup walkthrough. The watch provides projects, chats, voice prompts, task state, and limits; the Mac remains the machine actually running Codex.
A safer control model
A small companion interface should begin read-only: status, elapsed time, current task, recent output, and alerts. Low-risk inputs can come next, such as answering an agent’s explicit question. Starting new work or approving commands that modify systems should require stronger confirmation and clear context.
Remote bridges should use authenticated pairing, encrypted transport, short-lived credentials, rate limits, revocation, and redacted logs. Avoid exposing a general shell. Show exactly which machine and project will receive an instruction, and keep an emergency stop available. Voice input also needs a confirmation screen because transcription errors can change intent.
Explore agent tools and implementation evidence in the AtlasRepo catalog.
What is the right minimum interface for a long-running agent? Would you prefer a watch, phone, chat bot, web dashboard, or notifications with approval links?