The queue grew on every day it ran, including the day it closed 831 tickets

Seven days of unattended agents. 3,040 tickets filed, 2,203 closed, and the open count went up on every single one of those days. Not one net-negative day. The best day closed 831 and still finished +23.

I went looking for what the 837 survivors were about. Sorted by what they reference: 312 the ticket queue itself, 180 the dashboard, 178 the release tool, 114 the agent supervisor, 77 the flag service. Five reference the blog. One references the public log this post is going to.

274 of the 820 open items — 33% — cite no file and no line number. They are assertions about ticket state: this ticket is merged but still reads approved, that slot marker names a session that ended. I spot-checked one. Its subject had been closed and reconciled by hand two days after the ticket about it was filed. That whole tier decays. It is a photograph of a queue, filed into the queue.

Of the 2,203 closures, 2,002 recorded no reason at all. Of the 201 that did, duplicate (91) outnumbers shipped-or-fixed (75). One defect was filed six times in 26 hours by three different workers. A deduplicating pass shipped on the 18th; on the 20th, six fresh captures of a single artifact were filed.

The default driving this is one line and it looked free when it was written: capture is a verb any worker can call, it costs nothing, and it needs no permission. Everything else a worker can do requires an approved slice, a free claim slot, and a gate that a human stands at. So when the external queue is dry, the cheapest legal action available to an idle agent is to observe the system and file what it saw — and the system is always observable, because that is what a health surface is for. The supply never runs out. It scales with worker count, and it does not converge.

My own contribution: 319 of those closures happened inside one twenty-minute window with no reason recorded on any of them. That was me clearing the board by hand. It worked, and it means that for 91% of everything this fleet has ever closed, I can no longer tell you whether it was fixed or was the fourth copy of something three workers filed the same afternoon.

The infrastructure is all still up — dashboard on :7777, flags on :7780, gitea, the CI runner. Only the timers are off. They have been off for six days. Tickets filed in that time: zero.