The overnight run is not the achievement

Someone asked a builder this week how long they let an agent run unattended. The answer was a number pulled off a dashboard, six point seven hours, next to a claim of averaging half a billion tokens a day. They also joke that the two metrics that matter in 2026 are tokens per day and grams of protein per day, which is funnier than it has any right to be. The same week, a coding tool reported that cloud agents now account for fifty six percent of its merged pull requests, up from one in ten in December.

Good numbers. They are also the wrong ones to copy.

Asked which tasks actually fit a long unattended run, that same builder gave a far less quotable answer. The ones where you already have a clear plan. You talk it through with the model, it hands back a plan, you read the plan, and only then do you let it go for hours. The unattended part is downstream of a document a human sat and read.

The vault’s notes on background automation say the same thing in flatter language. A 24/7 agent is not an agent that keeps thinking. It is a scheduled or event triggered workflow with a named owner, durable state, observable receipts, bounded permissions, and a policy for what happens when a run is missed, repeated, or overlapping. The reliable part is supposed to be boring infrastructure. Use the model where judgment is needed, not as the scheduler of record.

Put those together and the overnight run stops sounding like a flex. What is being claimed, really, is that the planning happened in the evening.

I will grant the other side its best point. Some tasks only show their shape once something starts running, and a long unattended pass is a cheap way to find that out. Fine. Do it on a copy, with receipts you can actually read in the morning.

The number worth reporting is not how many hours it ran while you slept. It is how much of the next morning went into undoing it.