RBB LAB: cap overnight agent runs at what you can review the next morning
A small studio on what changed when agents started running for hours unattended: every mid-task judgement moves into the brief, and the morning review queue, not launch capacity, sets how many runs are worth starting.
In The overnight PR: when agents work while you sleep, the RBB LAB studio describes the routine many teams now run: a task goes in at six in the evening and a pull request with four hours of unwatched work is waiting at nine the next morning. Their point is that this is real and useful, and that it moves the hard part to two places: the brief written before the run and the review done after it.
The first shift is that supervision disappears. When you watch an agent, you steer it constantly without counting it as work: away from a deprecated file, off an abstraction you don't want, past a question you answer in ten seconds. Unattended, those calls get made by the model, and a wrong assumption at hour one carries through the next three. Their brief now has five parts: the goal as an outcome, a checkable definition of done (usually a command that must pass), boundaries, known traps, and what to do when the spec runs out. The last one, explicit permission to stop and write down the question instead of guessing, changed their results more than anything else.
The second shift is the morning. One overnight PR is a pleasure; five is a queue, and the reviewer has none of the context the agent built up. Their rule is to cap unattended runs at what one person can genuinely review the next day, two or three for them, not six. They learned it the hard way: the week they launched too many parallel runs because they were cheap to start, they merged less than in a normal week, because everything sat behind review.
Why it matters for code review
This complicates the case for always-on agents without rejecting it. Overnight momentum only compounds if it lands as work someone can verify, and the binding constraint is the reviewer's fresh attention the next morning, not how many runs the harness can launch. A backlog of unverified diffs is, in their words, worse than no work at all. The useful posture shift is small: read an overnight diff as the work of a capable stranger who never saw the codebase and had no one to ask, not as a colleague's PR.