How to run coding agents while you are away from the desk: async supervision

The promise of an agent is that it works while you do something else. The reality, for most people, is that they watch it. Supervision that requires presence cancels the parallelism it was supposed to create. Running agents while you are away from the desk is a different mode of working, and it has rules.

What is the switching tax?

Rubinstein, Meyer and Evans' task-switching experiments found that alternating between tasks carried a measurable time cost on every switch, larger for complex or unfamiliar tasks. The American Psychological Association's summary of that line of work puts the productive-time loss from heavy switching at up to 40%.

Watching an agent is a switching machine: you look, it is fine, you look away, it asks, you look back. Each look is a switch, and each one is paid for twice, once when you leave your other work and once when you return to it. Three sessions watched this way are not three times the output; they are three times the switching.

Synchronous versus asynchronous supervision

There are two ways to supervise an agent, and they produce very different days.

Synchronous supervision is safe and no faster than doing the work yourself for anything short. Asynchronous is the mode in which agents pay for themselves, and it needs three things: the ask has to reach you, it has to be answerable in seconds, and unanswered asks have to fail safe.

PeakWatch does this from your own data. Your coding agent, on your wrist. Join the free beta →

What makes async supervision safe

• Deny on timeout. An unanswered request never becomes a yes. If you are in a meeting, the agent waits or works around it; it does not guess.
• Arguments in the request. You are approving a file path or a command, not a verb. "Write" tells you nothing; "Write to tests/auth.test.ts" tells you enough.
• A summary on demand. After an hour away you should be able to rebuild context in one screen, not by scrolling a transcript.
• A list ranked by who is blocked. Three parallel sessions must not become three parallel anxieties. The one waiting on you goes to the top; the ones working stay quiet.

Fewer, better-framed decisions is the same principle behind a work uniform: remove the decisions that do not matter so the ones that do get a clear head.

What Anthropic's Remote Control gives you

Since early 2026 Claude Code has shipped a first-party way to leave the desk. Run claude remote-control in a project, or /remote-control inside a session, and the session appears in the Claude app on your phone and at claude.ai/code. The session keeps running on your machine; the phone is a window into it. Permission prompts are forwarded, and the app can push a notification when the agent needs a decision.

Push notifications are the part most people miss. Under /config there are two toggles: Push when Claude decides, for a message when a long task finishes, and Push when actions required, for permission prompts and questions. Turn on the second one at least; without it the ask sits in the app until you happen to open it, which is synchronous supervision with extra steps.

Claude Code skips pushes while you are typing in the connected terminal, so you are not notified about a prompt you are already looking at.

It covers the first requirement, reaching you, and most of the second. What it does not change is that your computer has to stay on and the claude process has to keep running; close the laptop and the session goes offline until you bring it back. The setup, the plans it needs and its limits are covered in their own guide.

How many sessions can one person supervise?

Two or three comfortably, if the asks are ranked and cheap to answer. Beyond that the summaries do the heavy lifting and the limiting factor becomes review, not supervision: every session produces a diff you have to read eventually.

The practical setup is one session per independent task, each in its own working directory so they cannot edit the same files. Claude Code's server mode does this for you: claude remote-control --spawn worktree gives each session its own git worktree, and the server accepts up to 32 concurrent sessions by default. Scoping, checkpointing and permissions for each session are what make the parallel ones safe to leave.

Rank by what needs you, not by what is newest. A working session needs nothing; a finished one needs review when you have time; a blocked one needs a decision now. A list sorted in that order is readable in a glance, and a glance is all the attention you want to spend on it while you are doing something else.

The wrist as the supervision surface

PeakWatch is built for the asynchronous mode. Sessions are listed by what needs you; a blocked agent sits at the top with a countdown. Tap to approve, dictate the next prompt, or ask for a summary of what happened. Sessions you started at a keyboard are mirrored too, and you can take one over from the watch. It is in private beta, on Wear OS watches including Galaxy Watch, with a relay you run on your own machine. You are somewhere else. The work continues.

Hen Eytan · Founder, UPeak

Builds the UPeak apps and writes up the research behind them.