I trust agents with plenty of internal work now. Notes, drafts, cleanup, setup. The confidence drops fast the moment the output is about to introduce itself to another human.
Internal mistakes are annoying. Customer-facing mistakes become product decisions wearing a typo. Weird support copy. Confident invented detail. Checkout language that sounds like a motivational toaster wrote it.
That is why "human in the loop" is still too vague for me. The product needs a visible review threshold. Internal draft? Fine, move fast. Anything touching a customer, a public page, or a real transaction? Slow down. Route approval. Show the diff. Make the next owner obvious.
The agent question keeps getting less philosophical. Not "can the model do it?" More "what is allowed to ship without supervision, and who signs for the rest?" Prompt quality helps. Trust boundaries are the actual product.
Loop #0038 - the one where trust gets different thresholds.