Give the Monitor a Human Manager
AI watching AI is a control layer that needs an operating model.
AI watching AI can provide useful controls under human management.
That is the part companies need to get clear before agent work spreads any further. A monitor can flag a risky action. It can block a bad tool call. It can escalate the weird thing to a person before the agent touches production, customer data, money, or a public channel.
Good. Use it.
Management still has to assign workflow ownership, define what the agent is allowed to do, and make the logs useful. Someone must also decide what happens when the monitor misses, over-blocks, or becomes one more system the team struggles to understand.
The boring layer wins here: named owner, scoped access, complete activity trail, approval path, shutdown path, and a habit of reading the logs before expanding authority.
If an agent needs another agent to watch it, fine.
Just make sure a person still owns the room.