Running one agent is a conversation. Running several at once is closer to keeping watch. The prompting you already know still matters, but the harder skill is deciding where your attention goes next and giving it to the session that would stall without you. This page is a working method for that, grounded in the way Helmsman lays your sessions out.
What does supervising look like in practice?
It is a short loop you repeat: scan the grid, focus the cell that needs you, steer it, then step back.
Helmsman keeps every session on screen at once, up to twelve live terminals per deck and fifty sessions in all, so the whole fleet is one glance away. You read the shape of the grid rather than the words in it. Most cells are running; one or two are waiting. You bring a waiting one into focus, type the answer or the next instruction, and hand it back. Then your eyes return to the whole board. The work is not typing faster. It is choosing the right cell each time and not letting a stalled one sit.
How do you tell which session needs you?
Three states matter, and you can tell them apart without reading a line: output still flowing, a prompt waiting for an answer, and silence.
A cell with output still scrolling is working; leave it alone. A cell holding at a question, an approval prompt or a “do you want to proceed?”, needs a single keystroke from you before it can move. Silence is the ambiguous one: the session may have finished, or it may be genuinely stuck. Scanning for the waiting cells is the highest-value habit you can build, because a missed approval is expensive. An agent that hit a permission prompt can sit idle for twenty minutes or more while you work another cell, doing nothing at all, simply because nothing surfaced it. The whole point of seeing every session at once is that these quiet holds do not hide.
How should you split work between agents?
Give each session work it can own from start to finish. Independent tasks parallelize cleanly; intertwined edits do not.
Independent means different repositories, or clearly separated areas of the same one: the API layer in one session, the docs in another. Two sessions editing the same files at the same time will collide and overwrite each other, so keep tightly coupled edits inside a single session. Then brief each session fully up front. The more complete the instruction, the longer a session runs before it has to stop and ask you something, and that quiet stretch is exactly what lets you hold more of them at once. A thin prompt buys you a session that pings back every few minutes; a complete brief buys you one that runs on its own.
What is the honest limit?
Your review is the bottleneck, not the machine, and there is no number worth defending.
How many sessions one person can hold is personal; Ways to run multiple Claude Code sessions covers the range people report, and the spread is wide. What matters on this page is how you find your own ceiling. It depends on the work, on how separate the repositories are, and on how much you trust each session to run without a check-in. The reliable move is to start with two or three, get the scan, focus, steer, step-back rhythm comfortable, and add one more only when the current set feels easy to watch. Growth here is by feel, not by target.
When should you close a session?
Close a cell when it is done, stuck past a nudge, or heading the wrong way.
Done is the easy case: the work is finished, so free the slot for the next task. Stuck means you have
already focused the cell, answered its question, nudged it once, and it is still spinning or confused;
a fresh session with a cleaner brief usually beats fighting a tangled one. Wrong direction means the
plan has drifted from what you wanted, and rather than negotiating it back you close and respawn.
Closing and respawning is cheap: each cell is your own local claude CLI, so a new session starts in
seconds, pointed at the same project. Treat a close as routine, not a failure.
None of this replaces good prompting; it sits on top of it. Keep the session grid in view, lean on the focus and steer loop for each cell that needs you, and the fleet stays productive without any one agent going quiet on you.