Autonomy Ladder

The autonomy ladder is a scale for how much of a task’s control flow the model owns, with five rungs. Manual: you do the work, perhaps with a model open in another window. Suggested: the system proposes, you execute. Drafted: the model produces the artifact and a human approves before anything leaves. Supervised: the model acts on its own inside bounds, with a person watching and able to stop it. Autonomous: an agent that acts and stops on its own judgment.

Autonomy is paid for in testability: every step the model owns is a decision you cannot assert on before it runs, and every step you keep is one you can test, cache, and cost in advance. A workflow sits low on the ladder by construction, which is why it is the default. Choose the rung by what the step needs, independent of how capable an agent could be built.

The rung is chosen per step, not per system. One pipeline can hold a deterministic parse, a drafted email that waits for approval, and a supervised research loop, each on the rung its risk earns. A step climbs only when its reality forces it: the paths genuinely cannot be enumerated in advance, or the case branches on a reply you cannot foresee.