I Put 358 Things in a Human Review Queue. I Reviewed Five.

Guardian, my personal memory-and-task system, once put every agent-written document in my review queue. The queue grew to 358 items.

I reviewed five.

Nobody looked at the other 353. The queue kept growing anyway.

So I stopped treating every document as the same kind of decision.

An internal working document can exist without my reading it. A message going out in my name needs my judgment. The agent finishing a draft doesn't make that decision for me.

For outbound writing, I only need the queue to ask whether this says what I mean and whether I want it sent.

A proposed memory needs another kind of review. Guardian might infer a standing food preference from one meal. If that conclusion reaches a later model request, it can shape an answer as though I stated it.

I keep every derived memory pending and out of later requests until I review it. The inference can't shape an answer while it waits. I accept that cost because an unread queue shouldn't turn every guess into a fact.

The outbound review decides whether something may go out in my name. Memory review has a different job. It keeps an inference from shaping later answers until I accept it.

IMO, those belong in separate queues with separate decisions.

My reading time shouldn't be spent checking work the application can establish directly.

If an article-saving workflow reports success, the application can retrieve the stored body and compare it with the intended text. A match tells me the saved copy is intact. I still read it and decide whether I approve publication. Those are the reasons to put the verified copy in front of me.

If the readback fails, the article may still have saved. The system should say “Saved copy not verified” and hold the delivery until it can check again.

I can inspect whether an article says what I mean. Before it reaches me, the application can check that the saved body matches the draft. I shouldn't have to spend one of five reviews discovering that its body is empty.