Last sprint closed 34 stories in seven days. Every one of them needed me to be sitting in one specific chair. This week that stopped being true.

Sprint 31 logged 120 hours against a 78-hour capacity. That number is only possible because the hours run in parallel lanes rather than end to end — several agents working at once, one operator.

Which is exactly where the ceiling was. The lanes scale. The operator does not.

And the operator has to be at the desk, because that is where the machine is.

I had already built half the answer

I spent most of last sprint making the cockpit fit a phone.

BUD became an installable Android app. It got a master dashboard, then a bottom nav bar with an alerts tab. Analytics went in, both directions. Four stories, thirteen hours, all pointed at the same idea: I should be able to see the business from my pocket.

It worked. I could see everything.

I still could not touch anything.

Watching is not driving

The thing that changed this week is called Remote Control. I run claude remote-control in a project directory on my own machine, or /rc inside a session I have already started, and that session shows up on my phone.

Not a copy of it. Not a remote environment. The same local session.

The code still runs on my machine. The filesystem is still my filesystem. The phone is simply a window onto the real thing.

Why the cloud version would not have done this

My work is not in a repository. It is in an estate.

Three OCI vaults across three tenancies. A private mesh network between the servers. A GPU in a laptop in another room that cooks the feature images for these articles. Fifty-one live sites on a register that exists nowhere outside the estate. And the documentation repo itself, which is the thing every agent reads before it does anything at all.

A cloud sandbox gets the repository. It does not get any of the rest, and it never will, because the entire point of the estate is that it is mine and it is private.

Remote Control does not need to get any of it. It is already standing inside it.

The part I did not expect

The first time it mattered, I was not at the desk. I had set a long run going and walked away from it, the way you do when you have already decided the next hour belongs to something else.

The phone buzzed. It was not a notification about the work. It was the work — a permission prompt, waiting on me, carrying the same context I would have had in the terminal. I read it. I approved it. The run carried on without me going anywhere near the machine it was running on.

It is a smaller moment than I expected and it changed more than I expected.

When a turn runs long, the terminal now offers to hand it to my phone. After a few approvals in a row it offers again. And if the laptop sleeps or the network drops, the session queues the messages and the prompts, then delivers them when the machine comes back.

That last one is what makes this real rather than clever. Closing a lid is no longer an interruption. It is a pause.

The number this actually moves

The second article in this series named three KPIs for an AI OS. The first was Away From Desk Autonomy: nought per cent, heading for sixty.

I have been measuring it wrong for six months. I was counting how much of the work runs without me. The honest measure is how much of it survives me walking away from the desk — and until this week the answer was almost none of it, because every approval, every correction and every no, do it this way needed the chair.

Sixty per cent was never going to be reached by making the agents better. It was going to be reached by making the chair optional.