Your agents are acting as your people, right now.

A named agent, not a product tile

The platform is one thing; the agents built on it are another. Each named instance becomes its own reviewable resource, with its own identity, its own credentials and its own list of people it can act as.

The platform is not the agent

Copilot Studio is one record. The agents your teams built on it are records of their own, each with its own reach.

Type sticks once it is known

An agent that has been identified stays identified. A later scan cannot quietly demote it back into the app list.

Every agent has a runtime behind it

The service principal that actually executes is linked to the agent, so the thing you review and the thing that holds the token are connected.

Gone means archived, not deleted

An agent that disappears keeps its record, and an identity that stops backing it stays on the record as dormant. You can always answer what it used to reach.

Every action, a setting

Each tool Luna can use sits in one of four states, from auto-approve to blocked, and the default when nothing is configured is that she waits for you. Guarded sits between: auto-approve only when your conditions hold.

Four states, one row each

The four states are auto-approve, guarded, needs approval and blocked. Every action Luna can take is in the list, and nothing is hidden from it.

Guarded is a rule, not a mood

Auto-approve applies only while your conditions hold: the department, the identity type, the status. When nothing matches, it comes back to you.

The default is that she asks

An action with no setting requires confirmation. Silence is never read as permission.

Settings inherit, and say where from

Settings cascade from organisation to automation, chat and job, and every screen tells you which level the rule came from before you change it.

Know what your agents can be.