For the human collaboration part, what's the hand-off model? does the agent block waiting for a human to resolve a comment, or is it fire-and-forget?