The night shift

On the night of 24 October I gave my manager agent one last instruction: “As many levels as possible. Now I’m off.” At 01:26 there were five new game worlds, each built from a free asset pack, each measured at 60 frames a second on an older iPhone I had left plugged in, and a summary waiting for my morning.

I did not sit up for it. That is the point of this note.

Agents stop

Coding agents do not get tired, but they do stop. At the end of a task. At a question. At a permission prompt. Sometimes for no reason I can find. A stopped session waits until someone looks at it, and at night nobody does. Left alone, a session that stops at midnight waits until morning.

In an earlier note I described the manager: one Claude Code session that keeps the records for my side projects and coordinates the others. It is also what keeps a night going.

Before bed: the plan

A night starts in the evening, with a plan I agree with the manager: the items in order, what “done” means for each (tests, checks, a measurement), and the answers to the questions I can see coming. The plan also says what may not happen overnight: anything public, anything that costs money, anything that cannot be undone.

Then I say go, and go to bed.

During the night: the manager watches

• Every turn ends with a report. Each working session ends every turn with a message to the manager: what it did, what it checked, what is next.
• A stop is noticed. The manager asks to be told when a session goes idle. If the session stopped on a question the plan already answers, the manager gives it the answer and the session carries on.
• A question that is really mine is written down, not waited on. The session moves to the next item, and the question goes into the morning report.
• A fallback check every 30 minutes, in case a notice never comes.
• Routine steps I have delegated, the manager approves. A push to one of my private repositories, after the manager’s own checks, is one of them. Anything irreversible or public waits for me.
• The Mac stays awake, on power, with the app’s own keep-awake setting on.

In the morning: a report

What was done, what was checked and how, what was not done and why, and the questions only I can answer, each with the options. I read it in the morning and answer the questions; the work picks up from there.

Four nights, measured

• 19 to 20 October: the fixes from a game’s playtest. Every item merged in the planned order, the full test suite green after each merge (1,357 tests at the end), and every new test seen failing against a deliberate break first, more than 120 such checks. Done at 03:32. Nothing pushed or uploaded.
• 20 to 21 October: the first full draft of my guide: 23 commits, and seven demo days recorded in a virtual machine for its examples. The report came at about 00:55.
• 24 to 25 October: the five game worlds above, finished at 01:26.
• 28 to 29 October: a ten-step plan for the guide’s next version. All ten built and committed by about 02:27; the morning report’s questions answered by 08:05.

What each of them needed from me was the plan before and the report after.

The uncomfortable part

• Sessions do stall. Not with an error: they just sit and wait. One did exactly that on 28 October while I was at the gym. That is why the manager watches, and why there is a fallback check behind the idle notice.
• Things break with nobody there. On 23 October a review stopped overnight on a network error, and was resumed the next morning.
• The last step is still mine. On 19 October I held back an upload to my testers that could have gone out overnight, until my own check in the morning. Irreversible steps wait for a person, and that person sleeps.
• “Around the clock” has limits. Usage limits apply, so a night runs as far as the plan and the subscription allow. And the Mac has to stay on.
• It is not “no human decisions”. The manager makes only the decisions the plan has already made. A good night needs a good plan, and the plan is my work, done before bed.

The method behind the manager is written up as a guide with a working kit, One Manager, Many Agents: on Gumroad, with more about it on its page here.