Operating modes describe who leads a piece of work once its decisions have been assigned autonomy levels: the coworker, the coworker with a person approving selected decisions, or the person with agents expanding what they can do. Autonomy levels are set decision by decision; the operating mode is what those decisions add up to for the process as a whole.
Why the mode is a design question
Most conversations about AI adoption ask which tasks can use AI. The more useful question is where humans should lead, where machines should lead, and where machines can operate independently. Getting that allocation wrong in either direction is expensive: a process forced to run autonomously before its decisions are reliable produces failures, and an expert forced to approve every trivial action produces a bottleneck with a dashboard.
The mode is not fixed per profession or department. The same employee may operate across all three in a day, and the same workflow may contain all three at once.
The three modes
Fully agentic — the coworker leads and runs
It gathers context, makes the decisions, takes the actions, and handles routine exceptions. People set boundaries, watch performance, and intervene when it meets conditions outside its authority. Appropriate where the underlying decisions are reliable, observable, and low-risk to delegate.
Human-in-the-loop — the coworker leads; a person approves
The work happens in the background, and selected decisions surface for approval, clarification, or final judgment in the queue or inbox the person already uses. Likely the dominant mode for operational AI. The question is never whether a human is in the loop — it is which decisions need one, and why.
Agent-in-the-loop — the person leads; agents extend them
For high-judgment, exploratory, or research-heavy work where the correct path cannot be specified in advance. Agents research, compare, prepare options, and execute bounded pieces; the expert decides what matters. A lawyer with agents can search, read, and synthesize far more — the consequential judgments stay with the lawyer.
Moving between modes
A process often starts agent-in-the-loop, becomes human-in-the-loop as the coworker builds a record, and may eventually run fully agentic in parts. That progression is driven by evidence, not enthusiasm, and it is not inevitable — for some work, human-in-the-loop is the permanent optimum, and for others the human is the primary source of value and should stay in the lead. Fully agentic is not the superior end state. The right mode is the one that produces the best outcome given the judgment required, the consequences of error, and the economics of human involvement.