Ambient status vs. remote control: two ways to watch an agent
Every mobile app for AI coding agents is making one of two bets. Knowing which one you want makes the choice between them obvious.
There are now a dozen or so iOS apps aimed at people running AI coding agents. They look similar in the App Store and they are not similar at all. Nearly all of them are making one of two distinct bets.
Bet one: remote control
Happy Coder, Omnara, Paseo and Codex Relay all take this side. The premise: your agent is running on your laptop, you are not at your laptop, therefore you should be able to drive it from your phone. Read the output, approve the diff, send the next prompt, keep it moving.
It is a real need and these are real products — Happy Coder charts in the top twenty grossing Developer Tools apps, which is not something you fake. If you genuinely want to work while away from your desk, this is the category you want.
Bet two: ambient status
Codeaton and Beacon take the other side. The premise: you do not want to work from your phone. You want to stop thinking about the agent entirely, and be pulled back only when it finishes or gets stuck.
Under this bet, the ideal number of times you open the app is zero. Everything of value happens on the Lock Screen, in a widget, or in the Island. The app is where you go to configure it, not where you go to use it.
How to tell which one you want
A question that sorts it cleanly: when your agent finishes something while you are away from your desk, what do you want to happen?
- If the answer is “I want to review it and kick off the next thing” — you want remote control.
- If the answer is “I want to know, and deal with it when I sit back down” — you want ambient status.
Most people find they want the second one more often than they expect, and the first one more urgently when they want it at all. Which is a good argument for status being the default surface and control being the thing you reach for deliberately.
The thing both bets share
Both exist because agent runs got long enough to be worth leaving. That is a genuinely new condition. Two years ago the model answered in seconds and there was nothing to monitor; now a task can run for forty minutes and stop halfway to ask you a question, and the cost of not knowing is real. Whichever bet you prefer, that is the problem underneath both of them.