Definition
The capacity to act with consequence on an institution's behalf, held by a machine, a human, or an assembly of both. Delegated agency is created by an authority grant and exercised through a decision system; it is not a property of the model that happens to be running.
It matters because agency is the quantity that accumulates. An institution that tracks capability but not delegated agency cannot say how much of its consequential action it has handed over, or whether it still holds the correction capacity to take any of it back. Agency handed over under a revisable delegation can be taken back one grant at a time; agency handed over without one comes back only by rebuilding what replaced it.
Tests that show it is in place
- The institution can list, for one workflow, every action class a machine may take on its behalf.
- Each listed action class traces to a named grant holder rather than to a team or a system name.
- Growth in delegated action classes over a quarter is reported alongside growth in correction capacity.
Evidence to keep
- Artifact: A register of action classes a machine may perform on the institution's behalf, with holders.
- Behavior: A new action class appearing only after an authorization decision, visible in the register history.
- Metric: Count of authorized action classes per workflow, tracked against correction capacity over time.
References
- Laws for engineering delegated intelligence · Doctrine