“Human in the loop” is not an architecture.
It is a phrase. Sometimes a useful one, sometimes an excuse to postpone the real design work.
The architectural questions are more concrete: Which decisions may the system make? At what confidence? Under which mission intent? What information must be observable? When does authority return to a human? What happens if the link to that human is gone?
Intent is the compression layer.
Remote systems cannot depend on continuous detailed instruction and still be meaningfully autonomous. They need a compact representation of mission intent: objectives, boundaries, priorities and rules for uncertainty.
That intent is what allows local decision-making without turning autonomy into uncontrolled behaviour.
Intervention must survive degraded communications.
A beautiful emergency-stop button in a control room is irrelevant if the communications architecture cannot support it when the environment becomes difficult. Intervention paths have to be designed with the same seriousness as autonomous functions.
Sometimes intervention means a direct command. Sometimes it means a pre-agreed safe mode. Sometimes the correct response is to continue locally because stopping would create more risk.
Explainability is operational.
The interesting question is not whether every internal model state can be translated into a paragraph. The interesting question is whether an operator can understand why the system believes a decision is safe, what assumptions are currently weak and what alternative actions remain available.
Autonomy without intent is just distance between a decision and the person who will be asked to explain it.
Responsibility does not disappear.
Automation redistributes work. Autonomy redistributes decision-making. Neither redistributes responsibility into thin air. The architecture has to make that responsibility explicit — technically, operationally and institutionally.
Continue the conversation