Sometimes the agent hits a blocker while the person who could help is unavailable. That is manageable when you have one, two, or five agents, because someone can step in, inspect the problem, and unblock the work manually.
The setup changes once dozens (ok, scale it - dozens of dozens) of agents are working across several projects at the same time. At that scale, coordination needs to exist inside the system itself: an agent should recognise that it is blocked, record the reason, and escalate the request to another agent that has the right permissions, context, or authority to resolve it. Human in the loop still matters, but it should not become a single point of failure. A good agent system knows when to ask for help, who can provide it, and how to continue the task once the blocker is removed.
