Earned autonomy
Earned autonomy is the process of allowing a specific kind of agent action to run automatically only after the record shows it has been reviewed, tested, and approved reliably.
Earned autonomy means an agent does not receive freedom because the system seems clever.
It earns a specific freedom through a record. A certain kind of reply, update, or routine action may begin held for human approval. If the work is approved unchanged over time, passes its tests, and stays inside written limits, that specific action can graduate.
The freedom is narrow. It does not transfer to other responsibilities.
Why Autonomy Should Be Reversible
Autonomy should be easier to lose than to gain.
If tests slip, complaints appear, or the context changes, the system should pull the freedom back and hold the action for review again. The point is not maximum automation. The point is appropriate trust.