IMO the biggest issue with CLIs for agents is to know when the agent is allowed to type. When is the command fully proceed, and next tokens can now be generated.