The problem
A prompt describes desired behaviour, but does not establish who is acting, which resources may be used, for how long or under whose responsibility. Without identity there is no dependable authorisation, selective revocation or understandable audit.
The architectural principle
Every agent needs a distinct identity, constrained credentials, explicit permissions and an operational owner. Autonomy must be bounded by policy enforced outside the model, not entrusted to the model’s ability to follow instructions.
The operating model
High-impact actions require approval or separation of duties. Actions, decisions, tools and data must be reconstructable. Revocation must stop access and activity without relying on the agent’s cooperation.