Supervised Autonomy
What it means: The agent acts immediately, but all decisions are logged and surfaced to a human reviewer within a defined time window.
Use this when:
Speed matters but mistakes are recoverable
The cost of a wrong decision is moderate
A human review within hours or a day is sufficient to catch and correct errors
Examples: Prior authorization submissions, scheduling optimization, supply chain reordering.
The test: If we catch a mistake within 24 hours, is that fast enough?
How to implement: Insert a Checkpoint command immediately before the agent would complete the action.
Human-in-the-Loop
What it means: The agent recommends, but a human approves before any action is taken.
Use this when:
The decision directly affects a person's wellbeing or safety
A wrong action cannot be easily undone
Accountability must clearly rest with a human
Examples: Clinical decision support, treatment protocol modifications, anything touching patient care.
The test: Does this decision affect a patient's health or safety?
How to implement: Insert a pauseForHumanIntervention command immediately before the agent would complete the action.
Note: if no approval is received within a 30 minute window, the action does not proceed and the run will end.