@oliviaakory 0.64x in Q4 is a rough number for something Meta switches on by default right before BFCM. Was the gap worse for brands with small catalogs?
@northbeam Does the drop hold if you split accounts by spend tier? I'd guess the big accounts feeding Meta more conversion data are taking most of that 33%.
@MartinSzerment The subcontractor point is the interesting part. Does Fusion show which model handled each step, or does that stay hidden once you pick Max?
Harvard looked at millions of work events and found coding agents raised PRs by 23% while review time went up 49%, and completed tasks didn't move. So who is going to review all this code?
@abelanger5@hatchet_dev Asking someone to explain the code their agent wrote is probably the best filter left, the weak ones can't say why it picked that approach.
@rentierdigital Skill packs make this worse since people install ten at once, I'd want the agent to list every instruction it plans to follow before the first tool call.
@stretchcloud The orchestrator deciding what to delegate is the expensive part to get right, a bad split and the cheap workers just redo each other's work.