Every night, fourteen agents are scheduled against my businesses. They start at 00:30 and the last one goes at 01:55. By the time I wake up there are drafts, a briefing, and a grade on the previous night’s work.
Here is last night, unedited, from the run logs:
00:30 doktrin-pflege no run — doctrine complete, valid, well-formed
00:30 strategie-impuls no run — already ran Tue 01.09
00:40 querdenker no run — 1 proposal undecided
00:50 diagnose no run — 1 diagnosis undecided
01:00 analytics ran
01:00 seo-blog ran
01:10 pr-zyklus ran
01:10 trend-briefing no run — already ran Sat 05.09
01:20 linkedin no run — already ran Sat 05.09
01:20 community-reddit no run — already ran Wed 02.09
01:30 community-x ran — draft, graded 5.5, rejected
01:40 report-kommentar ran
01:45 impuls ran
01:55 sales-inbox no run — already ran Fri 04.09Seven of fourteen did nothing. That is not a failure rate. It is the design.
Two kinds of gate
A cadence gate asks: has this already succeeded recently enough? A daily cron does not mean daily work. It means the agent gets a daily chance to notice that it is due. Most nights it is not.
A decision gate asks: is there an undecided thing of my kind still waiting for the human? Two agents stopped on that last night. The contrarian will not generate a new proposal while an old one sits unanswered, and the diagnostic will not run a new measurement while the previous finding is unread.
This one took me a while to accept. It feels wasteful: the machine could have produced something and did not. But an agent that generates faster than a human decides does not increase throughput. It builds a queue, and a queue nobody reads is indistinguishable from a machine that produces nothing — except that it costs money.
The grader
At 01:40 an agent reads what the others produced and grades it. Last night it gave the X draft a 5.5 and rejected it, with this reasoning:
The core claim says the same thing as two earlier drafts, just with a new frame. Cleanly done on the discretion side — no engine names, no CTA, product-free — but the thought is the third serving of the same number.
It was right. I would have posted it.
That grader had been switched on for two weeks and had graded exactly nothing, because it only runs against output that is marked as publishable, and the one loop it was attached to never marked anything. Two hundred and eighty artefacts, zero grades. A switch that is on and connected to nothing looks identical to a switch that is working.
Nothing goes out
None of these agents can publish, post, send, schedule, open an account or spend money. They produce drafts and stop. The last step is a human pressing a button, and the human is me.
That is where the bottleneck sits, and I would rather have it visible at the end of the chain than hidden in the middle of it.
