Definition
Plain language
Splitting authority across roles so no one component can do everything alone.
As stated in the literature
In evidence-carrying agent design, the architectural principle that the model proposing actions cannot also certify the facts authorizing them; evidence must come from independent verifiers.
Why it matters: Without this separation, an AI agent can talk itself into any action it wants by also being the one that grants permission; splitting roles is a core safety lever for autonomous systems.
For example, the agent that proposes paying out a refund isn't allowed to also confirm the customer is eligible — a separate verifier must check the records.