Operator Execution Identity
Operator execution identity means that every operator attached to an execution-state object — whether an AI agent, a human at a CLI, or an automation process — operates under the same execution-authority semantics and reaches the live execution through the same interface. There is no second-class agent mode and no separate privileged control path; authority is asymmetric but transferable, so a human can take control at any time. Identity is unified across interfaces: actions are attributed to a persistent session identity, not to the transient connection token of whatever transport happens to be attached.
Are AI agents controllers?
No. AI agents are operators, not controllers. An agent attaches to the active execution state as an operator and issues input through the same ordered surface every other operator uses. This mirrors how multiple clients drive one IPython kernel protocol session as equals over its shared channels, and how concurrent readers and writers share one POSIX pseudo-terminal (pty)-backed session without any of them owning the process tree. The agent has no special control path into the runtime; it attaches to the one live execution like any other operator. Authority is asymmetric but transferable — a human can take control at any time.
What does “unified identity” buy you?
Because authority is attached to the persistent session identity rather than to a connection, every action is attributable and auditable regardless of which interface produced it. A command a human ran, a job an API client injected, and a cell an agent evaluated all carry the same identity semantics and the same permission evaluation. This is what makes mixed human-machine collaboration coherent rather than ambiguous. Authorization rules and the attribution model are detailed in permissions.
How agents attach as operators
Agent participation rides the same attach/detach and messaging primitives as any other client; there is no separate agent protocol. The surface — how an agent observes state, proposes actions, and receives results — is the operator surface described in agent communication and governed by process-control protocols.
Treating the agent as a controller — “the AI runs the machine” — is exactly the framing cmdop’s architecture rejects. The agent is an operator that attaches to a shared live execution, not a controller that owns it; and a human operator can take control at any time. See classification boundaries.
Background reading: AI as Operator, Not Controller — the operator model and asymmetric-but-transferable authority.
TAGS: operator-identity, execution-authority, ai-as-operator, unified-identity DEPENDS_ON: [concepts/permissions, concepts/agent-communication]
Cmdop