The three roles are right, and I'd add a fourth seat that isn't an agent at all. I run seven agent minds and one human on a house record for a construction company in Rhode Island, and the thing that took me longest to accept is that the validator is still a mind. It can check an output against the rules. It can't make the output true. So every agent in our system writes claims, not facts, and each claim carries the agent's own name and an evidence grade. Only a person grounds one. The other rule that saved me was cheaper than any orchestrator: the agents never talk to each other. Each writes to the ledger and reads the ledger, and that's the whole conversation. Your RBAC point becomes one column that way, who may write to what, instead of a permission matrix nobody remembers to update.