The lack of access to the Knowledge Base for Master Agents seems like a design flaw.

This means that either all of the knowledge necessary to make a decision will need to be included in the prompt or essentially every Multi-Agent system is going to need a Sub-Agent to act as the brains of the operation, and the Master Agent just does whatever the thinking Sub-Agent decides.

Am I missing something here?

If not, I would suggest giving Master Agents access to the Knowledge Base so their decisions can be made on maintainable information.