I'm asking this because I've seen two camps:
Camp 1/ People who say they'd never hand over budget control.
Camp 2/ People already running fully automated spend with zero daily oversight.
The gap between these camps seem to be less about the technology and more about who gets blamed when it goes wrong.
I don't just mean debugging, I mean an actual account of the decision that holds up to someone else asking about it weeks later.
This is the problem as I see it: the path only existed at runtime and the model picked it, so there's nothing to go back to. Tool call logs tell you what happened, but not why that instead of something else.
Asking the model to explain itself afterwards gives you a plausible story rather than a record.
So either autonomy stays narrow enough that nobody asks, or there's an approach I haven't come across.
What are people actually running? I'm coming at this from a spend-heavy corner of the industry where this is becoming the blocker, but I suspect it's not specific to us.
I don't just mean debugging, I mean an actual account of the decision that holds up to someone else asking about it weeks later.
This is the problem as I see it: the path only existed at runtime and the model picked it, so there's nothing to go back to. Tool call logs tell you what happened, but not why that instead of something else.
Asking the model to explain itself afterwards gives you a plausible story rather than a record.
So either autonomy stays narrow enough that nobody asks, or there's an approach I haven't come across.
What are people actually running? I'm coming at this from a spend-heavy corner of the industry where this is becoming the blocker, but I suspect it's not specific to us.